Engage!

88 files changed, 24335 insertions, 0 deletions

diff --git a/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt b/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt
new file mode 100644
index 0000000..b6094fb
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt
@@ -0,0 +1,146 @@
+find_package(OpenCL REQUIRED)
+find_package(Python3 REQUIRED)
+
+set(TARGET_NAME ggml-opencl)
+
+ggml_add_backend_library(${TARGET_NAME}
+                         ggml-opencl.cpp
+                         ../../include/ggml-opencl.h)
+target_link_libraries(${TARGET_NAME} PRIVATE ${OpenCL_LIBRARIES})
+target_include_directories(${TARGET_NAME} PRIVATE ${OpenCL_INCLUDE_DIRS})
+
+if (GGML_OPENCL_PROFILING)
+    message(STATUS "OpenCL profiling enabled (increases CPU overhead)")
+    add_compile_definitions(GGML_OPENCL_PROFILING)
+endif ()
+
+add_compile_definitions(GGML_OPENCL_SOA_Q)
+add_compile_definitions(GGML_OPENCL_TARGET_VERSION=${GGML_OPENCL_TARGET_VERSION})
+
+if (GGML_OPENCL_USE_ADRENO_KERNELS)
+    message(STATUS "OpenCL will use matmul kernels optimized for Adreno")
+    add_compile_definitions(GGML_OPENCL_USE_ADRENO_KERNELS)
+endif ()
+
+if (GGML_OPENCL_EMBED_KERNELS)
+    add_compile_definitions(GGML_OPENCL_EMBED_KERNELS)
+
+    set(EMBED_KERNEL_SCRIPT "${CMAKE_CURRENT_SOURCE_DIR}/kernels/embed_kernel.py")
+    file(MAKE_DIRECTORY     "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
+
+    target_include_directories(${TARGET_NAME} PRIVATE "${CMAKE_CURRENT_BINARY_DIR}/autogenerated")
+endif ()
+
+function(ggml_opencl_add_kernel KNAME)
+    set(KERN_HDR ${CMAKE_CURRENT_BINARY_DIR}/autogenerated/${KNAME}.cl.h)
+    set(KERN_SRC ${CMAKE_CURRENT_SOURCE_DIR}/kernels/${KNAME}.cl)
+
+    if (GGML_OPENCL_EMBED_KERNELS)
+        message(STATUS "opencl: embedding kernel ${KNAME}")
+
+        # Python must be accessible from command line
+        add_custom_command(
+            OUTPUT ${KERN_HDR}
+            COMMAND ${Python3_EXECUTABLE} ${EMBED_KERNEL_SCRIPT} ${KERN_SRC} ${KERN_HDR}
+            DEPENDS ${KERN_SRC} ${EMBED_KERNEL_SCRIPT}
+            COMMENT "Generate ${KERN_HDR}"
+        )
+
+        target_sources(${TARGET_NAME} PRIVATE ${KERN_HDR})
+    else ()
+        message(STATUS "opencl: adding kernel ${KNAME}")
+        configure_file(${KERN_SRC} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${KNAME}.cl COPYONLY)
+    endif ()
+endfunction()
+
+set(GGML_OPENCL_KERNELS
+    add
+    add_id
+    argsort
+    tri
+    fill
+    clamp
+    cpy
+    cvt
+    diag_mask_inf
+    div
+    gelu
+    gemv_noshuffle_general
+    gemv_noshuffle
+    get_rows
+    glu
+    group_norm
+    solve_tri
+    im2col_f32
+    im2col_f16
+    mean
+    mul_mat_Ab_Bi_8x4
+    mul_mv_f16_f16
+    mul_mv_f16_f32_1row
+    mul_mv_f16_f32_l4
+    mul_mv_f16_f32
+    mul_mv_f32_f32
+    mul_mv_q4_0_f32
+    mul_mv_q4_0_f32_v
+    mul_mv_q4_0_f32_8x_flat
+    mul_mv_q4_0_f32_1d_8x_flat
+    mul_mv_q4_0_f32_1d_16x_flat
+    mul_mv_q4_k_f32
+    mul_mv_q6_k_f32
+    mul_mv_q6_k_f32_flat
+    mul_mv_q8_0_f32
+    mul_mv_q8_0_f32_flat
+    mul_mv_mxfp4_f32
+    mul_mv_mxfp4_f32_flat
+    mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_q8_0_f32
+    mul_mv_id_q8_0_f32_flat
+    mul_mv_id_mxfp4_f32
+    mul_mv_id_mxfp4_f32_flat
+    gemm_moe_mxfp4_f32
+    gemv_moe_mxfp4_f32
+    mul_mm_f32_f32_l4_lm
+    mul_mm_f16_f32_l4_lm
+    mul_mm_q8_0_f32_l4_lm
+    mul_mm_q6_k_f32_l4_lm
+    mul_mm_q8_0_f32_8x4
+    gemv_noshuffle_general_q8_0_f32
+    mul
+    norm
+    relu
+    rms_norm
+    rope
+    scale
+    set_rows
+    sigmoid
+    silu
+    softmax_4_f32
+    softmax_4_f16
+    softmax_f32
+    softmax_f16
+    sqr
+    sqrt
+    ssm_conv
+    sub
+    sum_rows
+    transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    expm1
+    softplus
+    pad
+    repeat
+    mul_mat_f16_f32
+    mul_mm_f16_f32_kq_kqv
+    conv2d
+    conv2d_f16_f32
+    flash_attn_f32_f16
+    flash_attn_f16
+    flash_attn_f32
+)
+
+foreach (K ${GGML_OPENCL_KERNELS})
+    ggml_opencl_add_kernel(${K})
+endforeach()
diff --git a/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp b/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp
new file mode 100644
index 0000000..40474c1
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp
@@ -0,0 +1,11165 @@
+#define CL_TARGET_OPENCL_VERSION GGML_OPENCL_TARGET_VERSION
+#define CL_USE_DEPRECATED_OPENCL_1_2_APIS
+
+// suppress warnings in CL headers for GCC and Clang
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#ifdef __clang__
+#pragma GCC diagnostic ignored "-Wgnu-anonymous-struct"
+#endif
+
+#include "ggml-opencl.h"
+#include "ggml-backend.h"
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+#include "ggml.h"
+
+#include <CL/cl.h>
+
+#include <inttypes.h>
+#include <string.h>
+
+#include <cstddef>
+#include <cstdint>
+#include <fstream>
+#include <vector>
+#include <string>
+#include <cmath>
+#include <map>
+#include <memory>
+#include <charconv>
+#include <mutex>
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
+
+#define UNUSED(x) (void)(x)
+
+#define CL_CHECK(err)                                               \
+    do {                                                            \
+        cl_int err_ = (err);                                        \
+        if (err_ != CL_SUCCESS) {                                   \
+            GGML_LOG_ERROR("ggml_opencl: %s error %d at %s:%d\n",  \
+                #err, err_, __FILE__, __LINE__);                    \
+            GGML_ASSERT(0);                                         \
+        }                                                           \
+    } while (0)
+
+//------------------------------------------------------------------------------
+// OpenCL
+//------------------------------------------------------------------------------
+
+bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor);
+
+// See https://gmplib.org/~tege/divcnst-pldi94.pdf figure 4.1.
+// Precompute mp (m' in the paper) and L such that division
+// can be computed using a multiply (high 32b of 64b result)
+// and a shift:
+//
+// n/d = (mulhi(n, mp) + n) >> L;
+struct fastdiv_vals {
+    uint32_t mp;
+    uint32_t L;
+    uint32_t d;
+    uint32_t pad;
+};
+static_assert(sizeof(fastdiv_vals) == 16, "fastdiv_vals size incorrect");
+
+static fastdiv_vals init_fastdiv_values(uint64_t d_64) {
+    GGML_ASSERT(d_64 != 0);
+    GGML_ASSERT(d_64 <= std::numeric_limits<uint32_t>::max());
+
+    uint32_t d = (uint32_t)d_64;
+
+    // compute L = ceil(log2(d));
+    uint32_t L = 0;
+    while (L < 32 && (uint32_t{ 1 } << L) < d) {
+        L++;
+    }
+
+    uint32_t mp = (uint32_t) ((uint64_t{ 1 } << 32) * ((uint64_t{ 1 } << L) - d) / d + 1);
+    // pack divisor as well to reduce error surface
+    return { mp, L, d, 0 };
+}
+
+enum GPU_FAMILY {
+    ADRENO,
+    INTEL,
+    UNKNOWN,
+};
+
+enum ADRENO_GPU_GEN {
+    ADRENO_UNKNOWN,
+    A7X,
+    A8X,
+    X1E,
+};
+
+enum ADRENO_CL_COMPILER_TYPE {
+    E031,
+    DX,
+};
+
+struct ggml_cl_version {
+    cl_uint major = 0;
+    cl_uint minor = 0;
+};
+
+
+struct ggml_cl_compiler_version {
+    ADRENO_CL_COMPILER_TYPE type;
+    int major = -1;
+    int minor = -1;
+    int patch = -1;
+
+    bool same(ADRENO_CL_COMPILER_TYPE t, int x, int y, int z) const {
+        return major == x && minor == y && patch == z && type == t;
+    }
+    bool newer_than(ADRENO_CL_COMPILER_TYPE t, int x, int y, int z) const {
+        return major*10000 + minor*100 + patch > x*10000 + y*100 + z && type == t;
+    }
+    bool newer_than_or_same(ADRENO_CL_COMPILER_TYPE t, int x, int y, int z) const {
+        return same(t, x, y, z) || newer_than(t, x, y, z);
+    }
+};
+
+static size_t align_to(size_t value, size_t to_alignment) {
+    GGML_ASSERT(to_alignment && "Invalid alignment (must be non-zero)");
+    GGML_ASSERT((to_alignment & (to_alignment - 1)) == 0 && "to_alignment must be power-of-two");
+
+    return ((value + to_alignment - 1) / to_alignment) * to_alignment;
+}
+
+
+// Parses a version string of form "XX.YY ". On an error returns ggml_cl_version with all zeroes.
+static ggml_cl_version parse_cl_version(std::string_view str) {
+    size_t major_str_begin = 0;
+    size_t major_str_end   = str.find(".", major_str_begin);
+    if (major_str_end == std::string::npos) {
+        return {};
+    }
+
+    size_t minor_str_begin = major_str_end + 1;
+    size_t minor_str_end   = str.find(" ", minor_str_begin);
+    if (minor_str_end == std::string::npos) {
+        return {};
+    }
+
+    cl_uint version_major;
+    if (std::from_chars(str.data() + major_str_begin, str.data() + major_str_end, version_major).ec != std::errc{}) {
+        return {};
+    }
+
+    cl_uint version_minor;
+    if (std::from_chars(str.data() + minor_str_begin, str.data() + minor_str_end, version_minor).ec != std::errc{}) {
+        return {};
+    }
+    return { version_major, version_minor };
+}
+
+// Returns OpenCL platform's version. On an error returns ggml_cl_version with all zeroes.
+static ggml_cl_version get_opencl_platform_version(cl_platform_id platform) {
+    size_t param_size;
+    CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, 0, nullptr, &param_size));
+    std::unique_ptr<char[]> param_storage(new char[param_size]);
+    CL_CHECK(clGetPlatformInfo(platform, CL_PLATFORM_VERSION, param_size, param_storage.get(), nullptr));
+
+    auto              param_value    = std::string_view(param_storage.get(), param_size);
+    const std::string version_prefix = "OpenCL ";  // Suffix: "XX.YY <platform-specific-info>"
+    if (param_value.find(version_prefix) != 0) {
+        return {};
+    }
+    param_value.remove_prefix(version_prefix.length());
+    return parse_cl_version(param_value);
+}
+
+// Return a version to use in OpenCL C compilation. On an error returns ggml_cl_version with all zeroes.
+static ggml_cl_version get_opencl_c_version(ggml_cl_version platform_version, cl_device_id device) {
+    size_t param_size;
+
+#if CL_TARGET_OPENCL_VERSION >= 300
+    if (platform_version.major >= 3) {
+        CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, 0, nullptr, &param_size));
+        if (!param_size) {
+            return {};
+        }
+
+        std::unique_ptr<cl_name_version[]> versions(new cl_name_version[param_size]);
+        CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_ALL_VERSIONS, param_size, versions.get(), nullptr));
+        unsigned versions_count = param_size / sizeof(cl_name_version);
+
+        cl_version version_max = 0;
+        for (unsigned i = 0; i < versions_count; i++) {
+            version_max = std::max<cl_version>(versions[i].version, version_max);
+        }
+
+        return { CL_VERSION_MAJOR(version_max), CL_VERSION_MINOR(version_max) };
+    }
+#else
+    GGML_UNUSED(platform_version);
+#endif  // CL_TARGET_OPENCL_VERSION >= 300
+
+    CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, 0, nullptr, &param_size));
+    if (!param_size) {
+        return {};
+    }
+
+    std::unique_ptr<char[]> param_storage(new char[param_size]);
+    CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_OPENCL_C_VERSION, param_size, param_storage.get(), nullptr));
+    auto param_value = std::string_view(param_storage.get(), param_size);
+
+    const std::string version_prefix = "OpenCL C ";  // Suffix: "XX.YY <platform-specific-info>"
+    if (param_value.find(version_prefix) != 0) {
+        return {};
+    }
+    param_value.remove_prefix(version_prefix.length());
+
+    return parse_cl_version(param_value);
+}
+
+static ADRENO_GPU_GEN get_adreno_gpu_gen(const char *device_name) {
+    if (strstr(device_name, "730") ||
+        strstr(device_name, "740") ||
+        strstr(device_name, "750")) {
+        return ADRENO_GPU_GEN::A7X;
+    }
+
+    if (strstr(device_name, "830") ||
+        strstr(device_name, "840")) {
+        return ADRENO_GPU_GEN::A8X;
+    }
+
+    if (strstr(device_name, "X1")) {
+        return ADRENO_GPU_GEN::X1E;
+    }
+
+    return ADRENO_GPU_GEN::ADRENO_UNKNOWN;
+}
+
+static ggml_cl_compiler_version get_adreno_cl_compiler_version(const char *driver_version) {
+    std::string driver_ver_str(driver_version);
+    ADRENO_CL_COMPILER_TYPE type = ADRENO_CL_COMPILER_TYPE::E031;
+    size_t compiler_ver_pos = driver_ver_str.find("E031");
+    size_t compiler_ver_len = 13;
+    size_t compiler_major_offset = 5;
+    size_t compiler_minor_offset = 8;
+    size_t compiler_patch_offset = 11;
+
+    if (compiler_ver_pos == std::string::npos) {
+        compiler_ver_pos = driver_ver_str.find("DX");
+        if (compiler_ver_pos == std::string::npos) {
+            return {};
+        }
+        type = ADRENO_CL_COMPILER_TYPE::DX;
+        compiler_ver_len = 11;
+        compiler_major_offset = 3;
+    }
+
+    std::string compiler_ver_str = driver_ver_str.substr(compiler_ver_pos, compiler_ver_len);
+    int major = std::atoi(compiler_ver_str.substr(compiler_major_offset, 2).c_str());
+    int minor = std::atoi(compiler_ver_str.substr(compiler_minor_offset, 2).c_str());
+    int patch = std::atoi(compiler_ver_str.substr(compiler_patch_offset, 2).c_str());
+    return { type, major, minor, patch };
+}
+
+// cl buffer wrapper
+struct ggml_cl_buffer {
+    cl_mem buffer;
+    size_t size;
+
+    ggml_cl_buffer()
+        : buffer(nullptr), size(0) {}
+
+    ~ggml_cl_buffer() {
+        if (buffer) {
+            CL_CHECK(clReleaseMemObject(buffer));
+        }
+    }
+
+    void allocate(cl_context context, size_t new_size) {
+        if (new_size > size) {
+            size = new_size;
+            if (buffer) {
+                CL_CHECK(clReleaseMemObject(buffer));
+            }
+            cl_int err;
+            CL_CHECK((buffer = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
+        }
+    }
+};
+
+// Profiling
+struct ProfilingInfo {
+    std::string op_name;
+    std::string kernel_name;
+
+    cl_kernel kernel;
+    cl_event evt;
+
+    cl_ulong cmd_queued;
+    cl_ulong cmd_submit;
+    cl_ulong cmd_start;
+    cl_ulong cmd_end;
+    cl_ulong overhead_start;
+    cl_ulong overhead_end;
+    // For the times below, see spec for clGetEventProfilingInfo
+    // The time kernel spent in cmd queue - SUBMIT - QUEUED
+    cl_ulong cmd_queued_duration_ns;
+    // The time kernel spent for submission - START - SUBMIT
+    cl_ulong cmd_submit_duration_ns;
+    // Kernel execution time in nanoseconds - END - START
+    cl_ulong cmd_duration_ns;
+    // The time for the kernel to complete - COMPLETE - END
+    cl_ulong cmd_complete_duration_ns;
+    // Total time to finish the kernel - COMPELTE - QUEUED
+    cl_ulong cmd_total_duration_ns;
+    // Global and local work sizes.
+    size_t global_size[3];
+    size_t local_size[3];
+    // Op output size.
+    size_t output_size[4];
+};
+
+static void populateProfilingInfo(
+        ProfilingInfo& info, cl_event evt, cl_kernel kernel, cl_uint work_dim,
+        size_t global_size[3], size_t local_size[3],
+        const ggml_tensor * tensor) {
+    info.op_name     = tensor->name;
+    info.kernel      = kernel;
+    info.evt         = evt;
+
+    // 0 means not specified, e.g., 2D workgroup, or NULL for driver to choose
+    info.local_size[0] = 0;
+    info.local_size[1] = 0;
+    info.local_size[2] = 0;
+
+    info.global_size[0] = 0;
+    info.global_size[1] = 0;
+    info.global_size[2] = 0;
+
+    if (local_size) {
+        for (cl_uint i = 0; i < work_dim; ++i) {
+            info.local_size[i] = local_size[i];
+        }
+    }
+
+    for (cl_uint i = 0; i < work_dim; ++i) {
+        info.global_size[i] = global_size[i];
+    }
+
+    info.output_size[0] = tensor->ne[0];
+    info.output_size[1] = tensor->ne[1];
+    info.output_size[2] = tensor->ne[2];
+    info.output_size[3] = tensor->ne[3];
+}
+
+struct ggml_backend_opencl_context;
+
+// backend device context
+struct ggml_backend_opencl_device_context {
+    cl_platform_id platform;
+    std::string platform_name;
+
+    cl_device_id   device;
+    std::string    device_name;
+    cl_device_type device_type;
+    std::string    device_version;
+
+    // Initialized by ggml_cl2_init().
+    ggml_backend_opencl_context * backend_ctx = nullptr;
+
+    // Initialized by ggml_backend_opencl_device_get_buffer_type()
+    ggml_backend_buffer_type buffer_type;
+
+    cl_context context = nullptr;
+};
+
+// backend context
+struct ggml_backend_opencl_context {
+    int ref_count;
+
+    cl_device_id device;
+    std::string device_name;
+
+    std::string driver_version;
+
+    GPU_FAMILY gpu_family;
+    ADRENO_GPU_GEN adreno_gen;
+
+    cl_int alignment;
+    size_t max_alloc_size;
+    size_t max_workgroup_size;
+    bool fp16_support;
+    bool has_vector_subgroup_broadcast;
+    bool disable_fusion;
+    ggml_cl_compiler_version adreno_cl_compiler_version;
+
+    int adreno_wave_size;
+
+    cl_bool non_uniform_workgroups;
+    size_t  image_max_buffer_size;
+
+    cl_context context;
+    cl_command_queue queue;
+
+    // prealloc buffers for transposing weights and activations
+    ggml_cl_buffer prealloc_quant_trans;
+    ggml_cl_buffer prealloc_scales_trans;
+    ggml_cl_buffer prealloc_act_trans;
+
+    // prealloc buffers for src0 and src1
+    ggml_cl_buffer prealloc_src0;
+    ggml_cl_buffer prealloc_src1;
+
+    cl_program program_add;
+    cl_program program_add_id;
+    cl_program program_clamp;
+    cl_program program_cpy;
+    cl_program program_cvt;
+    cl_program program_diag_mask_inf;
+    cl_program program_gelu;
+    cl_program program_gemv_noshuffle_general;
+    cl_program program_gemv_noshuffle;
+    cl_program program_get_rows;
+    cl_program program_set_rows;
+    cl_program program_glu;
+    cl_program program_im2col_f16;
+    cl_program program_im2col_f32;
+    cl_program program_mul_mat_Ab_Bi_8x4;
+    cl_program program_mul_mv_q4_0_f32;
+    cl_program program_mul_mv_q4_0_f32_v;
+    cl_program program_mul_mv_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
+    cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
+    cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_q8_0_f32, program_mul_mv_q8_0_f32_flat;
+    cl_program program_mul_mv_mxfp4_f32;
+    cl_program program_mul_mv_mxfp4_f32_flat;
+    cl_program program_mul_mv_f16_f16;
+    cl_program program_mul_mv_f16_f32_1row;
+    cl_program program_mul_mv_f16_f32_l4;
+    cl_program program_mul_mv_f16_f32;
+    cl_program program_mul_mv_f32_f32;
+    cl_program program_mul;
+    cl_program program_mul_mat_f16_f32_tiled;
+    cl_program program_mul_mm_f16_f32_kqv;
+    cl_program program_mul_mm_f16_f32_kq;
+    cl_program program_div;
+    cl_program program_sub;
+    cl_program program_norm;
+    cl_program program_relu;
+    cl_program program_rms_norm;
+    cl_program program_group_norm;
+    cl_program program_rope;
+    cl_program program_silu;
+    cl_program program_sigmoid;
+    cl_program program_softmax_f32;
+    cl_program program_softmax_f16;
+    cl_program program_softmax_4_f32;
+    cl_program program_softmax_4_f16;
+    cl_program program_argsort_f32_i32;
+    cl_program program_sum_rows_f32;
+    cl_program program_pad;
+    cl_program program_upscale;
+    cl_program program_conv_2d_f16;
+    cl_program program_conv_2d_f32;
+    cl_program program_conv_2d_f16_f32;
+    cl_program program_tsembd;
+    cl_program program_gemv_moe_mxfp4_f32, program_gemm_moe_mxfp4_f32;
+    cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_id_q8_0_f32, program_mul_mv_id_q8_0_f32_flat;
+    cl_program program_mul_mv_id_mxfp4_f32;
+    cl_program program_mul_mv_id_mxfp4_f32_flat;
+    cl_program program_mul_mm_f32_f32_l4_lm;
+    cl_program program_mul_mm_f16_f32_l4_lm;
+    cl_program program_mul_mm_q8_0_f32_l4_lm;
+
+    cl_kernel kernel_add, kernel_add_row, kernel_add_f16, kernel_add_row_f16;
+    cl_kernel kernel_mul, kernel_mul_row, kernel_mul_f16, kernel_mul_row_f16;
+    cl_kernel kernel_div, kernel_div_row, kernel_div_f16, kernel_div_row_f16;
+    cl_kernel kernel_sub, kernel_sub_row, kernel_sub_f16, kernel_sub_row_f16;
+    cl_kernel kernel_add_id;
+    cl_kernel kernel_scale_f32, kernel_scale_f32_4;
+    cl_kernel kernel_sqr_cont_f32, kernel_sqr_cont_f32_4, kernel_sqr_cont_f16, kernel_sqr_cont_f16_4;
+    cl_kernel kernel_sqrt_cont_f32, kernel_sqrt_cont_f32_4, kernel_sqrt_cont_f16, kernel_sqrt_cont_f16_4;
+    cl_kernel kernel_mean_f32;
+    cl_kernel kernel_silu, kernel_silu_4;
+    cl_kernel kernel_gelu, kernel_gelu_4;
+    cl_kernel kernel_gelu_erf, kernel_gelu_erf_4;
+    cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
+    cl_kernel kernel_relu;
+    cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
+    cl_kernel kernel_tri;
+    cl_kernel kernel_fill;
+    cl_kernel kernel_clamp;
+    cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
+              kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
+    cl_kernel kernel_norm, kernel_norm_mul_add;
+    cl_kernel kernel_rms_norm, kernel_rms_norm_mul;
+    cl_kernel kernel_group_norm, kernel_group_norm_mul_add;
+    cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
+    cl_kernel kernel_soft_max, kernel_soft_max_4;
+    cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16_q1;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
+    cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
+    cl_kernel kernel_set_rows_f32_i64, kernel_set_rows_f32_i32, kernel_set_rows_f16_i64, kernel_set_rows_f16_i32;
+    cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
+    cl_kernel kernel_rope_multi_f32, kernel_rope_multi_f16, kernel_rope_vision_f32, kernel_rope_vision_f16;
+    cl_kernel kernel_cpy_f16_f16, kernel_cpy_f16_f32, kernel_cpy_f32_f16, kernel_cpy_f32_f32;
+    cl_kernel kernel_mul_mat_f32_f32;
+    cl_kernel kernel_mul_mat_f16_f16;
+    cl_kernel kernel_mul_mat_f16_f32_1row;
+    cl_kernel kernel_mul_mat_f16_f32;
+    cl_kernel kernel_mul_mat_f16_f32_l4;
+    cl_kernel kernel_mul_mat_f16_f32_tiled;
+    cl_kernel kernel_mul_mm_f16_f32_kqv;
+    cl_kernel kernel_mul_mm_f16_f32_kq;
+    cl_kernel kernel_mul_mat_q4_0_f32, kernel_mul_mat_q4_0_f32_v;
+    cl_kernel kernel_convert_block_q4_0, kernel_restore_block_q4_0;
+    cl_kernel kernel_convert_block_mxfp4, kernel_convert_block_mxfp4_trans, kernel_restore_block_mxfp4, kernel_restore_block_mxfp4_trans;
+    cl_kernel kernel_convert_block_q8_0, kernel_restore_block_q8_0, kernel_restore_block_q8_0_trans;
+    cl_kernel kernel_mul_mat_q4_0_f32_8x_flat;
+    cl_kernel kernel_convert_block_q4_0_noshuffle;
+    cl_kernel kernel_restore_block_q4_0_noshuffle;
+    cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
+    cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
+    cl_kernel kernel_mul_mv_q4_K_f32;
+    cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_q6_K_f32_flat;
+    cl_kernel kernel_mul_mv_mxfp4_f32, kernel_mul_mv_mxfp4_f32_flat;
+    cl_kernel kernel_mul_mv_q8_0_f32, kernel_mul_mv_q8_0_f32_flat;
+    cl_kernel kernel_solve_tri_f32;
+    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
+    cl_kernel kernel_argsort_f32_i32;
+    cl_kernel kernel_sum_rows_f32;
+    cl_kernel kernel_repeat_f32;
+    cl_kernel kernel_pad;
+    cl_kernel kernel_tanh_f32, kernel_tanh_f32_4, kernel_tanh_f32_nc;
+    cl_kernel kernel_tanh_f16, kernel_tanh_f16_4, kernel_tanh_f16_nc;
+    cl_kernel kernel_expm1_f32_nd;
+    cl_kernel kernel_expm1_f16_nd;
+    cl_kernel kernel_softplus_f32_nd;
+    cl_kernel kernel_softplus_f16_nd;
+    cl_kernel kernel_upscale;
+    cl_kernel kernel_upscale_bilinear;
+    cl_kernel kernel_concat_f32;
+    cl_kernel kernel_conv_2d_f16;
+    cl_kernel kernel_conv_2d_f32;
+    cl_kernel kernel_conv_2d_f16_f32;
+    cl_kernel kernel_ssm_conv_f32_f32, kernel_ssm_conv_f32_f32_4;
+    cl_kernel kernel_timestep_embedding;
+    cl_kernel kernel_gemv_moe_mxfp4_f32, kernel_gemm_moe_mxfp4_f32;
+    cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_q8_0_f32, kernel_mul_mv_id_q8_0_f32_flat;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32_flat;
+    cl_kernel kernel_mul_mm_f32_f32_l4_lm;
+    cl_kernel kernel_mul_mm_f16_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q8_0_f32_l4_lm;
+    cl_kernel kernel_mul_mm_q6_k_f32_l4_lm;
+
+    std::vector<ProfilingInfo> profiling_info;
+
+    void write_profiling_info() {
+        FILE * fperf = fopen("cl_profiling.csv", "w");
+        if (!fperf) {
+            GGML_LOG_ERROR("Failed to open cl_profiling.csv\n");
+            return;
+        }
+
+        // Populate profiling info
+        for (ProfilingInfo & info : profiling_info) {
+            cl_ulong cmd_queued;
+            cl_ulong cmd_submit;
+            cl_ulong cmd_start;
+            cl_ulong cmd_end;
+            cl_ulong cmd_complete;
+
+            CL_CHECK(clWaitForEvents(1, &info.evt));
+            CL_CHECK(clGetEventProfilingInfo(
+                info.evt, CL_PROFILING_COMMAND_QUEUED, sizeof(cl_ulong), &cmd_queued, NULL));
+            CL_CHECK(clGetEventProfilingInfo(
+                info.evt, CL_PROFILING_COMMAND_SUBMIT, sizeof(cl_ulong), &cmd_submit, NULL));
+            CL_CHECK(clGetEventProfilingInfo(
+                info.evt, CL_PROFILING_COMMAND_START, sizeof(cl_ulong), &cmd_start, NULL));
+            CL_CHECK(clGetEventProfilingInfo(
+                info.evt, CL_PROFILING_COMMAND_END, sizeof(cl_ulong), &cmd_end, NULL));
+            CL_CHECK(clGetEventProfilingInfo(
+                info.evt, CL_PROFILING_COMMAND_COMPLETE, sizeof(cl_ulong), &cmd_complete, NULL));
+            CL_CHECK(clReleaseEvent(info.evt));
+
+            char kernel_name[512];
+            CL_CHECK(clGetKernelInfo(info.kernel, CL_KERNEL_FUNCTION_NAME,
+                sizeof(kernel_name), kernel_name, NULL));
+            info.kernel_name = kernel_name;
+
+            info.cmd_queued = cmd_queued;
+            info.cmd_submit = cmd_submit;
+            info.cmd_start  = cmd_start;
+            info.cmd_end    = cmd_end;
+
+            info.cmd_queued_duration_ns     = cmd_submit    - cmd_queued;
+            info.cmd_submit_duration_ns     = cmd_start     - cmd_submit;
+            info.cmd_duration_ns            = cmd_end       - cmd_start;
+            info.cmd_complete_duration_ns   = cmd_complete  - cmd_end;
+            info.cmd_total_duration_ns      = cmd_complete  - cmd_queued;
+        }
+
+        // Dump a csv
+        fprintf(fperf, "op name, kernel name, exec duration (ms), global size, local size, output size\n");
+        for (const ProfilingInfo & info : profiling_info) {
+            fprintf(fperf, "%s,%s,%f,%zux%zux%zu,%zux%zux%zu,%zux%zux%zux%zu\n",
+                info.op_name.c_str(), info.kernel_name.c_str(),
+                info.cmd_duration_ns/1.e6f,
+                info.global_size[0], info.global_size[1], info.global_size[2],
+                info.local_size[0], info.local_size[1], info.local_size[2],
+                info.output_size[0], info.output_size[1], info.output_size[2], info.output_size[3]);
+        }
+        fclose(fperf);
+
+        // Dump a simple chrome trace
+        FILE* ftrace = fopen("cl_trace.json", "w");
+        if (!ftrace) {
+            GGML_LOG_ERROR("Failed to open cl_trace.json\n");
+            return;
+        }
+
+        fprintf(ftrace, "[\n");
+        for (const ProfilingInfo & info : profiling_info) {
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Host\"},\n",
+                info.kernel_name.c_str(), info.cmd_queued/1000);
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Host\"},\n",
+                info.kernel_name.c_str(), info.cmd_submit/1000);
+
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"B\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Device\"},\n",
+                info.kernel_name.c_str(), info.cmd_start/1000);
+            fprintf(ftrace, "{\"name\": \"%s\", \"cat\": \"OpenCL\", \"ph\": \"E\", \"ts\": %" PRIu64 ", \"pid\": \"\", \"tid\": \"Device\"},\n",
+                info.kernel_name.c_str(), info.cmd_end/1000);
+        }
+        fclose(ftrace);
+    }
+
+    size_t get_kernel_workgroup_size(cl_kernel kernel) const {
+        size_t workgroup_size = 0;
+        size_t ret_size = 0;
+        CL_CHECK(
+            clGetKernelWorkGroupInfo(kernel, device, CL_KERNEL_WORK_GROUP_SIZE,
+                sizeof(size_t), &workgroup_size, &ret_size));
+        GGML_ASSERT(sizeof(size_t) == ret_size);
+        return workgroup_size;
+    }
+
+    void enqueue_ndrange_kernel(cl_kernel kernel, cl_uint work_dim, size_t *global_work_size, size_t *local_work_size, const ggml_tensor * tensor) {
+#ifdef GGML_OPENCL_PROFILING
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+
+        profiling_info.emplace_back();
+        populateProfilingInfo(profiling_info.back(), evt, kernel, work_dim, global_work_size, local_work_size, tensor);
+#else
+        GGML_UNUSED(tensor);
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, work_dim, NULL, global_work_size, local_work_size, 0, NULL, NULL));
+#endif
+    }
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    // Transpose kernels
+    cl_program program_transpose;
+
+    cl_kernel kernel_transpose_32;
+    cl_kernel kernel_transpose_32_16;
+    cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_buf;
+    cl_kernel kernel_transpose_16_4x1;
+
+    // Gemm and Gemv related programs, kernels, etc
+    cl_program program_CL_gemm;
+    cl_program program_CL_gemv_general;
+    cl_program program_CL_gemv_4096_1_11008;
+    cl_program program_CL_gemv_4096_1_4096;
+    cl_program program_CL_gemv_11008_1_4096;
+    cl_program program_CL_gemv_32000_1_4096;
+    cl_kernel CL_mul_mat_Ab_Bi_8x4;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
+    cl_kernel CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
+    cl_kernel kernel_mul_mm_q8_0_f32_8x4;
+    cl_kernel CL_mul_mat_vec_q8_0_f32;
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    void free() {
+        ref_count--;
+        if (ref_count == 0) {
+#ifdef GGML_OPENCL_PROFILING
+            write_profiling_info();
+            profiling_info.clear();
+#endif
+        }
+    }
+};
+
+// All registered devices with a default device in the front.
+static std::vector<ggml_backend_device> g_ggml_backend_opencl_devices;
+
+inline std::string read_file(const std::string &path) {
+  std::ifstream ifs(path);
+  if (!ifs) {
+    return "";
+  }
+  std::string text;
+  ifs.seekg(0, std::ios::end);
+  text.resize(ifs.tellg());
+  ifs.seekg(0, std::ios::beg);
+  ifs.read(&text[0], text.size());
+  return text;
+}
+
+static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer, const std::string &compile_opts) {
+    cl_program p;
+    char *program_log;
+    size_t program_size;
+    size_t log_size;
+    int err;
+
+    program_size = strlen(program_buffer);
+
+    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
+    if(err < 0) {
+        GGML_LOG_ERROR("OpenCL error creating program");
+        exit(1);
+    }
+
+    err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
+    if(err < 0) {
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
+        program_log = (char*) malloc(log_size + 1);
+        program_log[log_size] = '\0';
+        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
+        GGML_LOG_ERROR("ggml_opencl: kernel compile error:\n\n%s\n", program_log);
+        free(program_log);
+        exit(1);
+    }
+
+    return p;
+}
+
+static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_version opencl_c_version) {
+    cl_int err;
+
+    // compiler options for general kernels
+    auto opencl_c_std =
+        std::string("CL") + std::to_string(opencl_c_version.major) + "." + std::to_string(opencl_c_version.minor);
+    std::string compile_opts = std::string("-cl-std=") + opencl_c_std +
+                               " -cl-mad-enable -cl-unsafe-math-optimizations"
+                               " -cl-finite-math-only -cl-fast-relaxed-math";
+
+    GGML_LOG_INFO("ggml_opencl: loading OpenCL kernels");
+
+    // add
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "add.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("add.cl");
+#endif
+        backend_ctx->program_add =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add         = clCreateKernel(backend_ctx->program_add, "kernel_add", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row     = clCreateKernel(backend_ctx->program_add, "kernel_add_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_f16     = clCreateKernel(backend_ctx->program_add, "kernel_add_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_add_row_f16 = clCreateKernel(backend_ctx->program_add, "kernel_add_row_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // add_id
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "add_id.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("add_id.cl");
+#endif
+        backend_ctx->program_add_id =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_add_id = clCreateKernel(backend_ctx->program_add_id, "kernel_add_id", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // tri
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tri.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tri.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_tri = clCreateKernel(prog, "kernel_tri_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // fill
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "fill.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("fill.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_fill = clCreateKernel(prog, "kernel_fill_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // clamp
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "clamp.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("clamp.cl");
+#endif
+        backend_ctx->program_clamp =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_clamp = clCreateKernel(backend_ctx->program_clamp, "kernel_clamp", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // cpy
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "cpy.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("cpy.cl");
+#endif
+        backend_ctx->program_cpy =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_cpy_f16_f16 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f16_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f16_f32 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f16_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f32_f16 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f32_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_cpy_f32_f32 = clCreateKernel(backend_ctx->program_cpy, "kernel_cpy_f32_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // cvt
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "cvt.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("cvt.cl");
+#endif
+        backend_ctx->program_cvt =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_0_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q4_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_mxfp4_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_mxfp4_trans", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_mxfp4_trans = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4_trans", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_mxfp4 = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_mxfp4", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q8_0  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q8_0_trans  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q8_0_trans", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // diag_mask_inf
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "diag_mask_inf.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("diag_mask_inf.cl");
+#endif
+        backend_ctx->program_diag_mask_inf =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_diag_mask_inf_8 = clCreateKernel(backend_ctx->program_diag_mask_inf, "kernel_diag_mask_inf_8", &err), err));
+        CL_CHECK((backend_ctx->kernel_diag_mask_inf   = clCreateKernel(backend_ctx->program_diag_mask_inf, "kernel_diag_mask_inf", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gelu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gelu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gelu.cl");
+#endif
+        backend_ctx->program_gelu =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gelu         = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_4       = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf     = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_erf_4   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_erf_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_quick   = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick", &err), err));
+        CL_CHECK((backend_ctx->kernel_gelu_quick_4 = clCreateKernel(backend_ctx->program_gelu, "kernel_gelu_quick_4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // glu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "glu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("glu.cl");
+#endif
+        backend_ctx->program_glu =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_geglu           = clCreateKernel(backend_ctx->program_glu, "kernel_geglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu           = clCreateKernel(backend_ctx->program_glu, "kernel_reglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu          = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_oai      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_oai", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick     = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_reglu_f16       = clCreateKernel(backend_ctx->program_glu, "kernel_reglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_swiglu_f16      = clCreateKernel(backend_ctx->program_glu, "kernel_swiglu_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_erf_f16   = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_erf_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_geglu_quick_f16 = clCreateKernel(backend_ctx->program_glu, "kernel_geglu_quick_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // get_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "get_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("get_rows.cl");
+#endif
+        backend_ctx->program_get_rows =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_get_rows_f32  = clCreateKernel(backend_ctx->program_get_rows, "kernel_get_rows_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_get_rows_f16  = clCreateKernel(backend_ctx->program_get_rows, "kernel_get_rows_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_get_rows_q4_0 = clCreateKernel(backend_ctx->program_get_rows, "kernel_get_rows_q4_0", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // solve_tri_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "solve_tri.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("solve_tri.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_solve_tri_f32 = clCreateKernel(prog, "kernel_solve_tri_f32", &err), err));
+        GGML_LOG_CONT(".");
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // im2col_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "im2col_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("im2col_f32.cl");
+#endif
+        backend_ctx->program_im2col_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_im2col_f32 = clCreateKernel(backend_ctx->program_im2col_f32, "kernel_im2col_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // im2col_f16
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "im2col_f16.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("im2col_f16.cl");
+#endif
+        backend_ctx->program_im2col_f16 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_im2col_f16 = clCreateKernel(backend_ctx->program_im2col_f16, "kernel_im2col_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_q4_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_q4_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_q4_0_f32, "kernel_mul_mat_q4_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_0_f32_v
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_0_f32_v.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_0_f32_v.cl");
+#endif
+        backend_ctx->program_mul_mv_q4_0_f32_v =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_q4_0_f32_v = clCreateKernel(backend_ctx->program_mul_mv_q4_0_f32_v, "kernel_mul_mat_q4_0_f32_v", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_q4_0_f32_8x_flat, "kernel_mul_mat_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_0_f32_1d_8x_flat
+    // This kernel does not compiler on Adreno cl compiler 38.01. Skip it for
+    // those compiler versions since it is anyway not used for Adreno.
+    if (backend_ctx->gpu_family != ADRENO ||
+        backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 38, 11, 0) ||
+        backend_ctx->adreno_cl_compiler_version.type == DX) {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_0_f32_1d_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_0_f32_1d_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q4_0_f32_1d_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_q4_0_f32_1d_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_q4_0_f32_1d_8x_flat, "kernel_mul_mat_q4_0_f32_1d_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_0_f32_1d_16x_flat
+    // This kernel does not compiler on Adreno cl compiler 38.01. Skip it for
+    // those compiler versions since it is anyway not used for Adreno.
+    if (backend_ctx->gpu_family != ADRENO ||
+        backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 38, 11, 0) ||
+    backend_ctx->adreno_cl_compiler_version.type == DX) {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_0_f32_1d_16x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_0_f32_1d_16x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q4_0_f32_1d_16x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_q4_0_f32_1d_16x_flat = clCreateKernel(backend_ctx->program_mul_mv_q4_0_f32_1d_16x_flat, "kernel_mul_mat_q4_0_f32_1d_16x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q4_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q4_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q4_k_f32.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q4_K_f32 = clCreateKernel(prog, "kernel_mul_mv_q4_K_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q6_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q6_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_q6_K =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32 = clCreateKernel(backend_ctx->program_mul_mv_q6_K, "kernel_mul_mv_q6_K_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q6_k_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q6_k_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q6_k_f32_flat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q6_K_f32_flat = clCreateKernel(prog, "kernel_mul_mv_q6_K_f32_flat", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32, "kernel_mul_mv_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_q8_0_f32_flat, "kernel_mul_mv_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32, "kernel_mul_mv_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32_flat, "kernel_mul_mv_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_f16_f16
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_f16_f16.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_f16_f16.cl");
+#endif
+        backend_ctx->program_mul_mv_f16_f16 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f16 = clCreateKernel(backend_ctx->program_mul_mv_f16_f16, "kernel_mul_mat_f16_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_f16_f32_1row
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_f16_f32_1row.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_f16_f32_1row.cl");
+#endif
+        backend_ctx->program_mul_mv_f16_f32_1row =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_1row = clCreateKernel(backend_ctx->program_mul_mv_f16_f32_1row, "kernel_mul_mat_f16_f32_1row", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_f16_f32_l4
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_f16_f32_l4.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_f16_f32_l4.cl");
+#endif
+        backend_ctx->program_mul_mv_f16_f32_l4 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_l4   = clCreateKernel(backend_ctx->program_mul_mv_f16_f32_l4, "kernel_mul_mat_f16_f32_l4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_f16_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_f16_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_f16_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_f16_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32 = clCreateKernel(backend_ctx->program_mul_mv_f16_f32, "kernel_mul_mat_f16_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_f32_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_f32_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_f32_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_f32_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f32_f32 = clCreateKernel(backend_ctx->program_mul_mv_f32_f32, "kernel_mul_mat_f32_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mat_f16_f32_tiled
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mat_f16_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mat_f16_f32.cl");
+#endif
+        backend_ctx->program_mul_mat_f16_f32_tiled =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mat_f16_f32_tiled = clCreateKernel(backend_ctx->program_mul_mat_f16_f32_tiled, "mul_mat_f16_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f32_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f32_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f32_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f32_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f32_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f32_f32_l4_lm, "kernel_mul_mm_f32_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f16_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f16_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f16_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_f16_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_l4_lm, "kernel_mul_mm_f16_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_q8_0_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q8_0_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q8_0_f32_l4_lm.cl");
+#endif
+        backend_ctx->program_mul_mm_q8_0_f32_l4_lm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q8_0_f32_l4_lm = clCreateKernel(backend_ctx->program_mul_mm_q8_0_f32_l4_lm, "kernel_mul_mm_q8_0_f32_l4_lm", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_q6_k_f32_l4_lm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_q6_k_f32_l4_lm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_q6_k_f32_l4_lm.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_q6_k_f32_l4_lm = clCreateKernel(prog, "kernel_mul_mm_q6_k_f32_l4_lm", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_f16_f32_kq_kqv
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mm_f16_f32_kq_kqv.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mm_f16_f32_kq_kqv.cl");
+#endif
+        backend_ctx->program_mul_mm_f16_f32_kqv =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts+" -DKQV ");
+        backend_ctx->program_mul_mm_f16_f32_kq =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_kqv = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_kqv, "mul_mm_f16_f32_kqv", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_mm_f16_f32_kq = clCreateKernel(backend_ctx->program_mul_mm_f16_f32_kq, "mul_mm_f16_f32_kq", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul.cl");
+#endif
+        backend_ctx->program_mul =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul         = clCreateKernel(backend_ctx->program_mul, "kernel_mul", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_f16     = clCreateKernel(backend_ctx->program_mul, "kernel_mul_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_mul_row_f16 = clCreateKernel(backend_ctx->program_mul, "kernel_mul_row_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // norm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "norm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("norm.cl");
+#endif
+        backend_ctx->program_norm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_norm         = clCreateKernel(backend_ctx->program_norm, "kernel_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_norm_mul_add = clCreateKernel(backend_ctx->program_norm, "kernel_norm_mul_add", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // relu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "relu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("relu.cl");
+#endif
+        backend_ctx->program_relu =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_relu = clCreateKernel(backend_ctx->program_relu, "kernel_relu", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // rms_norm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "rms_norm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("rms_norm.cl");
+#endif
+        backend_ctx->program_rms_norm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_rms_norm     = clCreateKernel(backend_ctx->program_rms_norm, "kernel_rms_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_rms_norm_mul = clCreateKernel(backend_ctx->program_rms_norm, "kernel_rms_norm_mul", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // rope
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "rope.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("rope.cl");
+#endif
+        backend_ctx->program_rope =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_rope_norm_f32   = clCreateKernel(backend_ctx->program_rope, "kernel_rope_norm_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_norm_f16   = clCreateKernel(backend_ctx->program_rope, "kernel_rope_norm_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_neox_f32   = clCreateKernel(backend_ctx->program_rope, "kernel_rope_neox_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_neox_f16   = clCreateKernel(backend_ctx->program_rope, "kernel_rope_neox_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_multi_f32  = clCreateKernel(backend_ctx->program_rope, "kernel_rope_multi_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_multi_f16  = clCreateKernel(backend_ctx->program_rope, "kernel_rope_multi_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_vision_f32 = clCreateKernel(backend_ctx->program_rope, "kernel_rope_vision_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_rope_vision_f16 = clCreateKernel(backend_ctx->program_rope, "kernel_rope_vision_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // scale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "scale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("scale.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_scale_f32   = clCreateKernel(prog, "kernel_scale_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_scale_f32_4 = clCreateKernel(prog, "kernel_scale_f32_4", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // silu
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "silu.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("silu.cl");
+#endif
+        backend_ctx->program_silu =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_silu   = clCreateKernel(backend_ctx->program_silu, "kernel_silu", &err), err));
+        CL_CHECK((backend_ctx->kernel_silu_4 = clCreateKernel(backend_ctx->program_silu, "kernel_silu_4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // softmax_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softmax_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softmax_f32.cl");
+#endif
+        backend_ctx->program_softmax_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_soft_max = clCreateKernel(backend_ctx->program_softmax_f32, "kernel_soft_max", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // softmax_f16
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softmax_f16.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softmax_f16.cl");
+#endif
+        backend_ctx->program_softmax_f16 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_soft_max_f16 = clCreateKernel(backend_ctx->program_softmax_f16, "kernel_soft_max_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // softmax_4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softmax_4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softmax_4_f32.cl");
+#endif
+        backend_ctx->program_softmax_4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_soft_max_4 = clCreateKernel(backend_ctx->program_softmax_4_f32, "kernel_soft_max_4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // softmax_4_f16
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softmax_4_f16.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softmax_4_f16.cl");
+#endif
+        backend_ctx->program_softmax_4_f16 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_soft_max_4_f16 = clCreateKernel(backend_ctx->program_softmax_4_f16, "kernel_soft_max_4_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // flash_attn
+    {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src_f16 {
+                    #include "flash_attn_f16.cl.h"
+                };
+                const std::string kernel_src_f32 {
+                    #include "flash_attn_f32.cl.h"
+                };
+                const std::string kernel_src_f32_f16 {
+                    #include "flash_attn_f32_f16.cl.h"
+                };
+        #else
+                const std::string kernel_src_f16 = read_file("flash_attn_f16.cl");
+                const std::string kernel_src_f32 = read_file("flash_attn_f32.cl");
+                const std::string kernel_src_f32_f16 = read_file("flash_attn_f32_f16.cl");
+        #endif
+
+        if (!kernel_src_f16.empty() && !kernel_src_f32.empty() && !kernel_src_f32_f16.empty()) {
+            const struct { int dk; int dv; int bm; int bn; } fa_dims[] = {
+                { 40,  40, 32, 32}, { 64,  64, 64, 64}, { 80,  80, 64, 32}, { 96,  96, 64, 32},
+                {112, 112, 32, 32}, {128, 128, 32, 32}, {192, 128, 16, 16},
+                {192, 192, 16, 16}, {256, 256, 16, 16},
+            };
+
+            for (size_t i = 0; i < sizeof(fa_dims)/sizeof(fa_dims[0]); ++i) {
+                const int dk = fa_dims[i].dk;
+                const int dv = fa_dims[i].dv;
+                const int bm = fa_dims[i].bm;
+                const int bn = fa_dims[i].bn;
+                std::string OPTS = compile_opts +
+                    " -D DK=" + std::to_string(dk) +
+                    " -D DV=" + std::to_string(dv) +
+                    " -D BLOCK_M=" + std::to_string(bm) +
+                    " -D BLOCK_N=" + std::to_string(bn);
+
+                cl_program prog_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16.c_str(), OPTS);
+                cl_kernel k_f16, k_f16_q1;
+                CL_CHECK((k_f16 = clCreateKernel(prog_f16, "flash_attn_f16", &err), err));
+                CL_CHECK((k_f16_q1 = clCreateKernel(prog_f16, "flash_attn_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f16[{dk, dv}] = k_f16;
+                backend_ctx->kernels_flash_attn_f16_q1[{dk, dv}] = k_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f16));
+
+                cl_program prog_f32 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32.c_str(), OPTS);
+                cl_kernel k_f32, k_f32_q1;
+                CL_CHECK((k_f32 = clCreateKernel(prog_f32, "flash_attn_f32", &err), err));
+                CL_CHECK((k_f32_q1 = clCreateKernel(prog_f32, "flash_attn_f32_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32[{dk, dv}] = k_f32;
+                backend_ctx->kernels_flash_attn_f32_q1[{dk, dv}] = k_f32_q1;
+                CL_CHECK(clReleaseProgram(prog_f32));
+
+                cl_program prog_f32_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32_f16.c_str(), OPTS);
+                cl_kernel k_f32_f16, k_f32_f16_q1;
+                CL_CHECK((k_f32_f16 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16", &err), err));
+                CL_CHECK((k_f32_f16_q1 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32_f16[{dk, dv}] = k_f32_f16;
+                backend_ctx->kernels_flash_attn_f32_f16_q1[{dk, dv}] = k_f32_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f32_f16));
+
+                backend_ctx->kernels_flash_attn_bm[{dk, dv}] = bm;
+                backend_ctx->kernels_flash_attn_bn[{dk, dv}] = bn;
+            }
+            GGML_LOG_CONT(".");
+        }
+    }
+
+    // argsort
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "argsort.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("argsort.cl");
+#endif
+        backend_ctx->program_argsort_f32_i32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_argsort_f32_i32 = clCreateKernel(backend_ctx->program_argsort_f32_i32, "kernel_argsort_f32_i32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // div
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "div.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("div.cl");
+#endif
+        std::string compile_opts = std::string("-cl-std=") + opencl_c_std +
+                               " -cl-mad-enable -cl-finite-math-only ";
+
+        backend_ctx->program_div =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_div         = clCreateKernel(backend_ctx->program_div, "kernel_div", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row     = clCreateKernel(backend_ctx->program_div, "kernel_div_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_f16     = clCreateKernel(backend_ctx->program_div, "kernel_div_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_div_row_f16 = clCreateKernel(backend_ctx->program_div, "kernel_div_row_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // sqr
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqr.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqr.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32     = clCreateKernel(prog, "kernel_sqr_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f32_4   = clCreateKernel(prog, "kernel_sqr_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16     = clCreateKernel(prog, "kernel_sqr_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqr_cont_f16_4   = clCreateKernel(prog, "kernel_sqr_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // sqrt
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sqrt.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sqrt.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32     = clCreateKernel(prog, "kernel_sqrt_cont_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f32_4   = clCreateKernel(prog, "kernel_sqrt_cont_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16     = clCreateKernel(prog, "kernel_sqrt_cont_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sqrt_cont_f16_4   = clCreateKernel(prog, "kernel_sqrt_cont_f16_4", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mean
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mean.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mean.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mean_f32 = clCreateKernel(prog, "kernel_mean_f32", &err), err));
+
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // sub
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sub.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sub.cl");
+#endif
+        backend_ctx->program_sub =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sub         = clCreateKernel(backend_ctx->program_sub, "kernel_sub", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_f16     = clCreateKernel(backend_ctx->program_sub, "kernel_sub_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_sub_row_f16 = clCreateKernel(backend_ctx->program_sub, "kernel_sub_row_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // sum_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sum_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sum_rows.cl");
+#endif
+        backend_ctx->program_sum_rows_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sum_rows_f32 = clCreateKernel(backend_ctx->program_sum_rows_f32, "kernel_sum_rows_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // sigmoid
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "sigmoid.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("sigmoid.cl");
+#endif
+        backend_ctx->program_sigmoid =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_sigmoid_f32 = clCreateKernel(backend_ctx->program_sigmoid, "kernel_sigmoid_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_sigmoid_f16 = clCreateKernel(backend_ctx->program_sigmoid, "kernel_sigmoid_f16", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // group_norm
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "group_norm.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("group_norm.cl");
+#endif
+        backend_ctx->program_group_norm =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_group_norm         = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm", &err), err));
+        CL_CHECK((backend_ctx->kernel_group_norm_mul_add = clCreateKernel(backend_ctx->program_group_norm, "kernel_group_norm_mul_add", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // repeat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "repeat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("repeat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_repeat_f32 = clCreateKernel(prog, "kernel_repeat_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // pad
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "pad.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("pad.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_pad =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_pad = clCreateKernel(backend_ctx->program_pad, "kernel_pad", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: pad kernel source not found or empty. Pad operations will not be available.\n");
+            backend_ctx->program_pad = nullptr;
+            backend_ctx->kernel_pad = nullptr;
+        }
+    }
+
+    // tanh
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tanh.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tanh.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_tanh_f32    = clCreateKernel(prog, "kernel_tanh_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f32_4  = clCreateKernel(prog, "kernel_tanh_f32_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f32_nc = clCreateKernel(prog, "kernel_tanh_f32_nc", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16    = clCreateKernel(prog, "kernel_tanh_f16", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16_4  = clCreateKernel(prog, "kernel_tanh_f16_4", &err), err));
+        CL_CHECK((backend_ctx->kernel_tanh_f16_nc = clCreateKernel(prog, "kernel_tanh_f16_nc", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // expm1
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "expm1.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("expm1.cl");
+#endif
+        cl_program prog;
+        if (!kernel_src.empty()) {
+            prog =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_expm1_f32_nd = clCreateKernel(prog, "kernel_expm1_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_expm1_f16_nd = clCreateKernel(prog, "kernel_expm1_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: expm1 kernel source not found or empty. Expm1 operation will not be available.\n");
+            prog = nullptr;
+            backend_ctx->kernel_expm1_f32_nd = nullptr;
+            backend_ctx->kernel_expm1_f16_nd = nullptr;
+        }
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // softplus
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softplus.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softplus.cl");
+#endif
+        cl_program prog;
+        if (!kernel_src.empty()) {
+            prog =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_softplus_f32_nd = clCreateKernel(prog, "kernel_softplus_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_softplus_f16_nd = clCreateKernel(prog, "kernel_softplus_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: softplus kernel source not found or empty. Softplus operation will not be available.\n");
+            prog = nullptr;
+            backend_ctx->kernel_softplus_f32_nd = nullptr;
+            backend_ctx->kernel_softplus_f16_nd = nullptr;
+        }
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // upscale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "upscale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("upscale.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_upscale =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_upscale = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale", &err), err));
+            if (backend_ctx->program_upscale) {
+                 cl_int err_bilinear;
+                 backend_ctx->kernel_upscale_bilinear = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale_bilinear", &err_bilinear);
+                 if (err_bilinear != CL_SUCCESS) {
+                    GGML_LOG_WARN("ggml_opencl: kernel_upscale_bilinear not found in upscale.cl. Bilinear upscale will not be available. Error: %d\n", err_bilinear);
+                    backend_ctx->kernel_upscale_bilinear = nullptr;
+                 }
+            } else {
+                backend_ctx->kernel_upscale_bilinear = nullptr;
+            }
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: upscale kernel source not found or empty. Upscale operations will not be available.\n");
+            backend_ctx->program_upscale = nullptr;
+            backend_ctx->kernel_upscale = nullptr;
+            backend_ctx->kernel_upscale_bilinear = nullptr;
+        }
+    }
+
+    // concat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "concat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("concat.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_concat_f32 = clCreateKernel(prog, "kernel_concat_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // timestep_embedding
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tsembd.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("tsembd.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tsembd =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_timestep_embedding = clCreateKernel(backend_ctx->program_tsembd, "kernel_timestep_embedding", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: timestep_embedding kernel source not found or empty. This op will not be available.\n");
+            backend_ctx->program_tsembd = nullptr;
+            backend_ctx->kernel_timestep_embedding = nullptr;
+        }
+    }
+
+    // set_rows
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "set_rows.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("set_rows.cl");
+#endif
+        backend_ctx->program_set_rows =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f32_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f32_i32", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i64 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i64", &err), err));
+        CL_CHECK((backend_ctx->kernel_set_rows_f16_i32 = clCreateKernel(backend_ctx->program_set_rows, "kernel_set_rows_f16_i32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+     // conv2d
+     {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src {
+                    #include "conv2d.cl.h"
+                };
+                const std::string kernel_src_f16_f32 {
+                    #include "conv2d_f16_f32.cl.h"
+                };
+        #else
+                const std::string kernel_src = read_file("conv2d.cl");
+                const std::string kernel_src_f16_f32 = read_file("conv2d_f16_f32.cl");
+        #endif
+                if (!kernel_src.empty()) {
+                    backend_ctx->program_conv_2d_f16 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), (std::string(compile_opts) + " -DUSE_FP16=1").c_str());
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16 = clCreateKernel(backend_ctx->program_conv_2d_f16, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                    backend_ctx->program_conv_2d_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f32 = clCreateKernel(backend_ctx->program_conv_2d_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16 = nullptr;
+                    backend_ctx->program_conv_2d_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f32 = nullptr;
+                }
+                if (!kernel_src_f16_f32.empty()) {
+                    backend_ctx->program_conv_2d_f16_f32 =
+                        build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16_f32.c_str(), compile_opts);
+                    CL_CHECK((backend_ctx->kernel_conv_2d_f16_f32 = clCreateKernel(backend_ctx->program_conv_2d_f16_f32, "kernel_conv_2d", &err), err));
+                    GGML_LOG_CONT(".");
+                } else {
+                    GGML_LOG_WARN("ggml_opencl: conv2d_f16_f32 kernel source not found or empty. This op will not be available.\n");
+                    backend_ctx->program_conv_2d_f16_f32 = nullptr;
+                    backend_ctx->kernel_conv_2d_f16_f32 = nullptr;
+                }
+    }
+
+    // ssm_conv
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "ssm_conv.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("ssm_conv.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32   = clCreateKernel(prog, "kernel_ssm_conv_f32_f32", &err), err));
+        CL_CHECK((backend_ctx->kernel_ssm_conv_f32_f32_4 = clCreateKernel(prog, "kernel_ssm_conv_f32_f32_4", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_q4_0_f32_8x_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q4_0_f32_8x_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q4_0_f32_8x_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q4_0_f32_8x_flat, "kernel_mul_mv_id_q4_0_f32_8x_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_q8_0_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32, "kernel_mul_mv_id_q8_0_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_q8_0_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_q8_0_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_q8_0_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_q8_0_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_q8_0_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_q8_0_f32_flat, "kernel_mul_mv_id_q8_0_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32, "kernel_mul_mv_id_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mv_id_mxfp4_f32_flat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32_flat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32_flat.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32_flat =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32_flat, "kernel_mul_mv_id_mxfp4_f32_flat", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // Adreno kernels
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    // transpose
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "transpose.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("transpose.cl");
+#endif
+        backend_ctx->program_transpose =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_buf = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_buf", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_general
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable "
+                                       " -DSIMDGROUP_WIDTH=" +
+                                       std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_CL_gemv_general {
+            #include "gemv_noshuffle_general.cl.h"
+        };
+#else
+        const std::string kernel_src_CL_gemv_general = read_file("gemv_noshuffle_general.cl");
+#endif
+
+        backend_ctx->program_CL_gemv_general = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv_general.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general = clCreateKernel(backend_ctx->program_CL_gemv_general, "kernel_gemv_noshuffle", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle
+    {
+        // Gemv 2048, 16384
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+            " -cl-mad-enable "
+            " -DLINE_STRIDE_A=2048 "
+            " -DBLOCK_STRIDE_A=16384 "
+            " -DSIMDGROUP_WIDTH=" +
+            std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_CL_gemv {
+            #include "gemv_noshuffle.cl.h"
+        };
+#else
+        const std::string kernel_src_CL_gemv = read_file("gemv_noshuffle.cl");
+#endif
+
+        backend_ctx->program_CL_gemv_4096_1_4096 = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_4096, "kernel_gemv_noshuffle", &err), err));
+        GGML_LOG_CONT(".");
+
+        // Gemv 2048, 16384
+        CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+            " -cl-mad-enable "
+            " -DLINE_STRIDE_A=2048 "
+            " -DBLOCK_STRIDE_A=16384 "
+            " -DSIMDGROUP_WIDTH=" +
+            std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+        backend_ctx->program_CL_gemv_4096_1_11008 = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008 = clCreateKernel(backend_ctx->program_CL_gemv_4096_1_11008, "kernel_gemv_noshuffle", &err), err));
+        GGML_LOG_CONT(".");
+
+        // Gemv 5504, 44032
+        CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+            " -cl-mad-enable "
+            " -DLINE_STRIDE_A=5504 "
+            " -DBLOCK_STRIDE_A=44032 "
+            " -DSIMDGROUP_WIDTH=" +
+            std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+        backend_ctx->program_CL_gemv_11008_1_4096 = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_11008_1_4096, "kernel_gemv_noshuffle", &err), err));
+        GGML_LOG_CONT(".");
+
+        // Gemv 16000, 128000
+        CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+            " -cl-mad-enable "
+            " -DLINE_STRIDE_A=16000 "
+            " -DBLOCK_STRIDE_A=128000 "
+            " -DSIMDGROUP_WIDTH=" +
+            std::to_string(backend_ctx->adreno_wave_size);
+
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+        backend_ctx->program_CL_gemv_32000_1_4096 = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv.c_str(), CL_gemv_compile_opts);
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096 = clCreateKernel(backend_ctx->program_CL_gemv_32000_1_4096, "kernel_gemv_noshuffle", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mat_Ab_Bi_8x4
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_CL_gemm {
+            #include "mul_mat_Ab_Bi_8x4.cl.h"
+        };
+#else
+        const std::string kernel_src_CL_gemm = read_file("mul_mat_Ab_Bi_8x4.cl");
+#endif
+        backend_ctx->program_CL_gemm = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_CL_gemm.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->CL_mul_mat_Ab_Bi_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mat_Ab_Bi_8x4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // mul_mm_q8_0_f32_8x4
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_q8_8x4_gemm {
+            #include "mul_mm_q8_0_f32_8x4.cl.h"
+       };
+#else
+        const std::string kernel_src_q8_8x4_gemm = read_file("mul_mm_q8_0_f32_8x4.cl");
+#endif
+        backend_ctx->program_CL_gemm = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_q8_8x4_gemm.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_mul_mm_q8_0_f32_8x4 = clCreateKernel(backend_ctx->program_CL_gemm, "kernel_mul_mm_q8_0_f32_8x4", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemv_noshuffle_general_q8_0_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable "
+                                       " -DSIMDGROUP_WIDTH=" +
+                                       std::to_string(backend_ctx->adreno_wave_size);
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAT ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src_CL_gemv_general {
+            #include "gemv_noshuffle_general_q8_0_f32.cl.h"
+        };
+#else
+        const std::string kernel_src_CL_gemv_general = read_file("gemv_noshuffle_general_q8_0_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src_CL_gemv_general.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->CL_mul_mat_vec_q8_0_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    std::string CL_moe_compile_opts = std::string("-cl-std=") + opencl_c_std +
+            " -cl-mad-enable "
+            " -cl-fast-relaxed-math";
+
+    // gemv_moe_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_moe_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_moe_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_gemv_moe_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemv_moe_mxfp4_f32, "kernel_gemv_moe_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemm_moe_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_moe_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemm_moe_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_gemm_moe_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_moe_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemm_moe_mxfp4_f32 = clCreateKernel(backend_ctx->program_gemm_moe_mxfp4_f32, "kernel_gemm_moe_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_LOG_CONT("\n");
+}
+
+// XXX static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
+// XXX    static bool initialized = false;
+// XXX    static ggml_backend_opencl_context *backend_ctx = nullptr;
+
+static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev);
+
+namespace /* anonymous */ {
+extern struct ggml_backend_device_i ggml_backend_opencl_device_i;
+}
+
+// Look for available and suitable devices.
+static std::vector<ggml_backend_device> ggml_opencl_probe_devices(ggml_backend_reg * reg) {
+    std::vector<ggml_backend_device> found_devices;
+
+#ifdef GGML_OPENCL_PROFILING
+    GGML_LOG_INFO("ggml_opencl: OpenCL profiling enabled\n");
+#endif
+
+    struct cl_device;
+    struct cl_platform {
+        cl_platform_id id;
+        unsigned number;
+        char name[128];
+        char vendor[128];
+        struct cl_device * devices;
+        unsigned n_devices;
+        struct cl_device * default_device;
+    };
+
+    struct cl_device {
+        struct cl_platform * platform;
+        cl_device_id id;
+        unsigned number;
+        cl_device_type type;
+        char name[128];
+        char version[128];
+    };
+
+    enum { NPLAT = 16, NDEV = 16 };
+
+    struct cl_platform platforms[NPLAT];
+    unsigned n_platforms = 0;
+    struct cl_device devices[NDEV];
+    unsigned n_devices = 0;
+    struct cl_device * default_device = NULL;
+    unsigned           default_platform_number = 0;
+
+    cl_platform_id platform_ids[NPLAT];
+    if (clGetPlatformIDs(NPLAT, platform_ids, &n_platforms) != CL_SUCCESS) {
+        GGML_LOG_ERROR("ggml_opencl: plaform IDs not available.\n");
+        return found_devices;
+    }
+
+    for (unsigned i = 0; i < n_platforms; i++) {
+        struct cl_platform * p = &platforms[i];
+        p->number = i;
+        p->id = platform_ids[i];
+        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
+        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
+
+        cl_device_id device_ids[NDEV];
+        cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
+        if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
+            p->n_devices = 0;
+        } else {
+            CL_CHECK(clGetDeviceIDsError);
+        }
+        p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
+        p->default_device = NULL;
+
+        for (unsigned j = 0; j < p->n_devices; j++) {
+            struct cl_device * d = &devices[n_devices];
+            d->number = n_devices++;
+            d->id = device_ids[j];
+            d->platform = p;
+            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
+            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
+            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_VERSION, sizeof(d->version), &d->version, NULL));
+
+            if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
+                p->default_device = d;
+            }
+        }
+
+        if (default_device == NULL && p->default_device != NULL) {
+            default_device          = p->default_device;
+            default_platform_number = i;
+        }
+    }
+
+    if (n_devices == 0) {
+        GGML_LOG_ERROR("ggml_opencl: could find any OpenCL devices.\n");
+        return found_devices;
+    }
+
+    char *      user_platform_string = getenv("GGML_OPENCL_PLATFORM");
+    char *      user_device_string   = getenv("GGML_OPENCL_DEVICE");
+    int         user_platform_number = -1;
+    int         user_device_number   = -1;
+    cl_device * candidate_devices    = nullptr;
+    unsigned    n_candidate_devices  = 0;
+
+    unsigned n;
+    if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
+        user_platform_number = (int)n;
+    }
+    if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
+        user_device_number = (int)n;
+    }
+    if (user_platform_number != -1 && user_device_number != -1) {
+        cl_platform* platform = &platforms[user_platform_number];
+        if ((unsigned)user_device_number >= platform->n_devices) {
+            GGML_LOG_ERROR("ggml_opencl: invalid device number %d\n", user_device_number);
+            exit(1);
+        }
+        default_device      = &platform->devices[user_device_number];
+        candidate_devices   = platform->devices;
+        n_candidate_devices = platform->n_devices;
+    } else {
+        // Choose a platform by matching a substring.
+        if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
+            for (unsigned i = 0; i < n_platforms; i++) {
+                struct cl_platform * p = &platforms[i];
+                if (strstr(p->name, user_platform_string) != NULL ||
+                    strstr(p->vendor, user_platform_string) != NULL) {
+                    user_platform_number = (int)i;
+                    break;
+                }
+            }
+            if (user_platform_number == -1) {
+                GGML_LOG_ERROR("ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
+                exit(1);
+            }
+        }
+
+        int                  platform_idx = user_platform_number != -1 ? user_platform_number : default_platform_number;
+        struct cl_platform * p            = &platforms[platform_idx];
+        candidate_devices                 = p->devices;
+        n_candidate_devices               = p->n_devices;
+        default_device                    = p->default_device;
+        if (n_candidate_devices == 0) {
+            GGML_LOG_ERROR("ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
+            exit(1);
+        }
+
+        if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
+            for (unsigned i = 0; i < n_candidate_devices; i++) {
+                struct cl_device * d = &candidate_devices[i];
+                if (strstr(d->name, user_device_string) != NULL) {
+                    user_device_number = d->number;
+                    break;
+                }
+            }
+            if (user_device_number == -1) {
+                GGML_LOG_ERROR("ggml_opencl: no device matching '%s' was found.\n", user_device_string);
+                exit(1);
+            }
+        }
+        if (user_device_number != -1) {
+            candidate_devices   = &devices[user_device_number];
+            n_candidate_devices = 1;
+            default_device      = &candidate_devices[0];
+        }
+
+        GGML_ASSERT(n_candidate_devices > 0);
+
+        if (default_device == NULL) {
+            default_device = &candidate_devices[0];
+        }
+    }
+
+    GGML_ASSERT(n_candidate_devices != 0 && candidate_devices);
+
+    // Put the default device in front.
+    for (unsigned i = 1; i < n_candidate_devices; i++) {
+        if (&candidate_devices[i] == default_device) {
+            std::swap(candidate_devices[0], candidate_devices[i]);
+            default_device = &candidate_devices[0];
+            break;
+        }
+    }
+
+    GGML_LOG_INFO("ggml_opencl: selected platform: '%s'\n", default_device->platform->name);
+
+    std::vector<cl_device_id> device_ids;
+    for (auto dev = candidate_devices, dev_end = candidate_devices + n_candidate_devices; dev != dev_end; dev++) {
+        device_ids.push_back(dev->id);
+    }
+
+    cl_int                err;
+    cl_context            shared_context;
+    cl_context_properties properties[] = { (intptr_t) CL_CONTEXT_PLATFORM, (intptr_t) default_device->platform->id, 0 };
+
+    CL_CHECK(
+        (shared_context = clCreateContext(properties, device_ids.size(), device_ids.data(), NULL, NULL, &err), err));
+
+    for (auto dev = candidate_devices, dev_end = candidate_devices + n_candidate_devices; dev != dev_end; dev++) {
+        GGML_LOG_INFO("\nggml_opencl: device: '%s (%s)'\n", dev->name, dev->version);
+
+        auto dev_ctx = std::unique_ptr<ggml_backend_opencl_device_context>(new ggml_backend_opencl_device_context{
+            /*.platform         =*/dev->platform->id,
+            /*.platform_nane    =*/dev->platform->name,
+            /*.device           =*/dev->id,
+            /*.device_name      =*/dev->name,
+            /*.device_type      =*/dev->type,
+            /*.device_version   =*/dev->version,
+            /*.backend_ctx      =*/nullptr,
+            /*.buffer_type      =*/{},
+            /*.context          =*/shared_context,
+        });
+
+        found_devices.push_back(ggml_backend_device{
+            /* .iface   = */ ggml_backend_opencl_device_i,
+            /* .reg     = */ reg,
+            /* .context = */ dev_ctx.get(),
+        });
+
+        if (!ggml_cl2_init(&found_devices.back())) {
+            found_devices.pop_back();
+            GGML_LOG_INFO("ggml_opencl: drop unsupported device.\n");
+            continue;
+        }
+
+        dev_ctx.release();
+    }
+
+    if (found_devices.size()) {
+        auto * dev_ctx = static_cast<ggml_backend_opencl_device_context *>(found_devices.front().context);
+        GGML_LOG_INFO("ggml_opencl: default device: '%s (%s)'\n", dev_ctx->device_name.c_str(),
+                      dev_ctx->device_version.c_str());
+
+        if (dev_ctx->device_type != CL_DEVICE_TYPE_GPU) {
+            GGML_LOG_WARN("ggml_opencl: warning, the default device is not a GPU: '%s'.\n",
+                          dev_ctx->device_name.c_str());
+        }
+    }
+
+    return found_devices;
+}
+
+// Initialize device if it is supported (returns nullptr if it is not).
+static ggml_backend_opencl_context * ggml_cl2_init(ggml_backend_dev_t dev) {
+    GGML_ASSERT(dev);
+    GGML_ASSERT(dev->context);
+
+    ggml_backend_opencl_device_context * dev_ctx = (ggml_backend_opencl_device_context *) dev->context;
+    GGML_ASSERT(dev_ctx->platform);
+    GGML_ASSERT(dev_ctx->device);
+
+    if (dev_ctx->backend_ctx) {
+        return dev_ctx->backend_ctx;
+    }
+
+    auto backend_ctx        = std::make_unique<ggml_backend_opencl_context>();
+    backend_ctx->device     = dev_ctx->device;
+    backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
+
+    // ref_count get increased in ggml_backend_opencl_device_init
+    // This function is also used to retrieve backend context, so we don't want
+    // to increase ref_count for each call. We only want to increase ref_count
+    // when the associated device is initialized
+    backend_ctx->ref_count  = 0;
+
+    if (strstr(dev_ctx->device_name.c_str(), "Adreno") ||
+        strstr(dev_ctx->device_name.c_str(), "Qualcomm") ||
+        strstr(dev_ctx->device_version.c_str(), "Adreno")) {
+        backend_ctx->gpu_family = GPU_FAMILY::ADRENO;
+        // Usually device version contains the detailed device name
+        backend_ctx->adreno_gen = get_adreno_gpu_gen(dev_ctx->device_version.c_str());
+        if (backend_ctx->adreno_gen == ADRENO_GPU_GEN::ADRENO_UNKNOWN) {
+            backend_ctx->adreno_gen = get_adreno_gpu_gen(dev_ctx->device_name.c_str());
+        }
+
+        // Use wave size of 64 for all Adreno GPUs.
+        backend_ctx->adreno_wave_size = 64;
+    } else if (strstr(dev_ctx->device_name.c_str(), "Intel")) {
+        backend_ctx->gpu_family = GPU_FAMILY::INTEL;
+    } else {
+        GGML_LOG_ERROR("Unsupported GPU: %s\n", dev_ctx->device_name.c_str());
+        backend_ctx->gpu_family = GPU_FAMILY::UNKNOWN;
+        return nullptr;
+    }
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    if (backend_ctx->gpu_family != GPU_FAMILY::ADRENO) {
+        GGML_LOG_ERROR("ggml_opencl: Adreno-specific kernels should not be enabled for non-Adreno GPUs; "
+            "run on an Adreno GPU or recompile with CMake option `-DGGML_OPENCL_USE_ADRENO_KERNELS=OFF`\n");
+        return nullptr;
+    }
+#endif
+
+    // Populate backend device name
+    backend_ctx->device_name = dev_ctx->device_name;
+
+    // A local ref of cl_device_id for convenience
+    cl_device_id device = backend_ctx->device;
+
+    ggml_cl_version platform_version = get_opencl_platform_version(dev_ctx->platform);
+
+    // Check device OpenCL version, OpenCL 2.0 or above is required
+    ggml_cl_version opencl_c_version = get_opencl_c_version(platform_version, device);
+    if (opencl_c_version.major < 2) {
+        GGML_LOG_ERROR("ggml_opencl: OpenCL 2.0 or above is required\n");
+        return nullptr;
+    }
+
+    // Check driver version
+    size_t driver_version_str_size;
+    clGetDeviceInfo(device, CL_DRIVER_VERSION, 0, NULL, &driver_version_str_size);
+    char *driver_version = (char *)alloca(driver_version_str_size + 1);
+    clGetDeviceInfo(device, CL_DRIVER_VERSION, driver_version_str_size, driver_version, NULL);
+    driver_version[driver_version_str_size] = '\0';
+    GGML_LOG_INFO("ggml_opencl: OpenCL driver: %s\n", driver_version);
+    backend_ctx->driver_version = driver_version;
+
+    backend_ctx->adreno_cl_compiler_version = get_adreno_cl_compiler_version(driver_version);
+    backend_ctx->has_vector_subgroup_broadcast =
+        (backend_ctx->adreno_cl_compiler_version.type == E031 && backend_ctx->adreno_cl_compiler_version.major >= 47) ||
+        (backend_ctx->adreno_cl_compiler_version.type == DX   && backend_ctx->adreno_cl_compiler_version.major >= 17);
+    GGML_LOG_INFO("ggml_opencl: vector subgroup broadcast support: %s\n",
+        backend_ctx->has_vector_subgroup_broadcast ? "true" : "false");
+
+    size_t ext_str_size;
+    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
+    char *ext_buffer = (char *)alloca(ext_str_size + 1);
+    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
+    ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
+    // Check if ext_buffer contains cl_khr_fp16
+    backend_ctx->fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
+    GGML_LOG_INFO("ggml_opencl: device FP16 support: %s\n", backend_ctx->fp16_support ? "true" : "false");
+
+    // fp16 is required
+    if (!backend_ctx->fp16_support) {
+        GGML_LOG_ERROR("ggml_opencl: device does not support FP16\n");
+        return nullptr;
+    }
+
+    // If OpenCL 3.0 is supported, then check for cl_khr_subgroups, which becomes
+    // optional in OpenCL 3.0 (cl_khr_subgroup is mandatory in OpenCL 2.x)
+    if (opencl_c_version.major == 3 && strstr(ext_buffer, "cl_khr_subgroups") == NULL &&
+        strstr(ext_buffer, "cl_intel_subgroups") == NULL) {
+        GGML_LOG_ERROR("ggml_opencl: device does not support subgroups (cl_khr_subgroups or cl_intel_subgroups) "
+            "(note that subgroups is an optional feature in OpenCL 3.0)\n");
+        return nullptr;
+    }
+
+    cl_uint base_align_in_bits;
+    CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_MEM_BASE_ADDR_ALIGN, sizeof(cl_uint), &base_align_in_bits, NULL));
+    GGML_ASSERT(base_align_in_bits % 8u == 0);
+    backend_ctx->alignment = base_align_in_bits / 8u;
+    GGML_LOG_INFO("ggml_opencl: mem base addr align: %u\n", backend_ctx->alignment);
+
+    clGetDeviceInfo(device, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof(size_t), &backend_ctx->max_alloc_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: max mem alloc size: %zu MB\n", backend_ctx->max_alloc_size/1024/1024);
+
+    clGetDeviceInfo(device, CL_DEVICE_IMAGE_MAX_BUFFER_SIZE, sizeof(size_t), &backend_ctx->image_max_buffer_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max image buffer size (pixels): %lu\n", backend_ctx->image_max_buffer_size);
+
+    clGetDeviceInfo(device, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof(size_t), &backend_ctx->max_workgroup_size, NULL);
+    GGML_LOG_INFO("ggml_opencl: device max workgroup size: %lu\n", backend_ctx->max_workgroup_size);
+
+    // Check SVM.
+    cl_device_svm_capabilities svm_caps;
+    CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_SVM_CAPABILITIES, sizeof(cl_device_svm_capabilities), &svm_caps, 0));
+    GGML_LOG_INFO("ggml_opencl: SVM coarse grain buffer support: %s\n",
+        svm_caps & CL_DEVICE_SVM_COARSE_GRAIN_BUFFER ? "true" : "false");
+    GGML_LOG_INFO("ggml_opencl: SVM fine grain buffer support: %s\n",
+        svm_caps & CL_DEVICE_SVM_FINE_GRAIN_BUFFER ? "true" : "false");
+    GGML_LOG_INFO("ggml_opencl: SVM fine grain system support: %s\n",
+        svm_caps & CL_DEVICE_SVM_FINE_GRAIN_SYSTEM ? "true" : "false");
+    GGML_LOG_INFO("ggml_opencl: SVM atomics support: %s\n",
+        svm_caps & CL_DEVICE_SVM_ATOMICS ? "true" : "false");
+
+    if (opencl_c_version.major >= 3) {
+        // Assume it is not available for 3.0, since it is optional in 3.0.
+        // If compiling against 3.0, then we can query.
+        backend_ctx->non_uniform_workgroups = false;
+#if CL_TARGET_OPENCL_VERSION >= 300
+        CL_CHECK(clGetDeviceInfo(device, CL_DEVICE_NON_UNIFORM_WORK_GROUP_SUPPORT, sizeof(cl_bool),
+                                 &backend_ctx->non_uniform_workgroups, 0));
+#endif
+    } else {
+        GGML_ASSERT(opencl_c_version.major == 2);
+        // Non-uniform workgroup sizes is mandatory feature in v2.x.
+        backend_ctx->non_uniform_workgroups = true;
+    }
+
+    // Print out configurations
+#ifdef GGML_OPENCL_SOA_Q
+    GGML_LOG_INFO("ggml_opencl: flattening quantized weights representation as struct of arrays (GGML_OPENCL_SOA_Q)\n");
+#endif // GGML_OPENCL_SOA_Q
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_LOG_INFO("ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)\n");
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    cl_int err;
+
+    // A local ref of cl_context for convenience
+    cl_context context = backend_ctx->context = dev_ctx->context;
+
+    //CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
+    //    (err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
+    //    (queue = clCreateCommandQueue(context, device, 0, &err), err)
+    //)));
+    cl_command_queue_properties command_queue_props = 0;
+#ifdef GGML_OPENCL_PROFILING
+    command_queue_props |= CL_QUEUE_PROFILING_ENABLE;
+#endif
+    CL_CHECK((backend_ctx->queue = clCreateCommandQueue(context, device, command_queue_props, &err), err));
+
+    // Load kernels
+    load_cl_kernels(backend_ctx.get(), opencl_c_version);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    // Allocate intermediate buffers and images
+    size_t required_A_q_d_bytes = 311164928;
+    size_t required_A_s_d_bytes = 38895616;
+    size_t required_B_d_bytes = 45088768;
+
+    // Ensure buffer sizes do not exceed the maximum allocation size
+    size_t max_A_q_d_bytes = MIN(required_A_q_d_bytes, backend_ctx->max_alloc_size);
+    size_t max_A_s_d_bytes = MIN(required_A_s_d_bytes, backend_ctx->max_alloc_size);
+    size_t max_B_d_bytes   = MIN(required_B_d_bytes, backend_ctx->max_alloc_size);
+    if (required_A_q_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: A_q_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_A_q_d_bytes, max_A_q_d_bytes);
+    }
+    if (required_A_s_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: A_s_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_A_s_d_bytes, max_A_s_d_bytes);
+    }
+    if (required_B_d_bytes > backend_ctx->max_alloc_size) {
+        GGML_LOG_WARN("ggml_opencl: B_d buffer size reduced from %zu to %zu due to device limitations.\n",
+                      required_B_d_bytes, max_B_d_bytes);
+    }
+
+    backend_ctx->prealloc_quant_trans.allocate(context, max_A_q_d_bytes);
+    backend_ctx->prealloc_scales_trans.allocate(context, max_A_s_d_bytes);
+    backend_ctx->prealloc_act_trans.allocate(context, max_B_d_bytes);
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    backend_ctx->disable_fusion = getenv("GGML_OPENCL_DISABLE_FUSION") != nullptr;
+
+    dev_ctx->backend_ctx = backend_ctx.release();
+    return dev_ctx->backend_ctx;
+}
+
+static void ggml_cl2_free(ggml_backend_t backend) {
+    ggml_backend_opencl_context * ctx = (ggml_backend_opencl_context *) backend->context;
+    ctx->free();
+
+    // The CL context is shared by all backends, release it if all backends have been released
+    bool should_release_opencl = true;
+    for (auto device : g_ggml_backend_opencl_devices) {
+        ggml_backend_opencl_device_context * ctx_dev = (ggml_backend_opencl_device_context *) device.context;
+        if (ctx_dev->backend_ctx->ref_count > 0) {
+            should_release_opencl = false;
+        }
+    }
+
+    if (should_release_opencl) {
+        CL_CHECK(clReleaseContext(ctx->context));
+    }
+}
+
+//------------------------------------------------------------------------------
+// Tensor extra management
+//------------------------------------------------------------------------------
+struct ggml_tensor_extra_cl {
+    // The buffer object that holds the data.
+    cl_mem data_device;
+    // The offset into the buffer object. This is primarily for scratch buffer
+    // and view operation.
+    // NB: this offset no longer includes view offset (view_offs). Whenever this
+    // offset is used, view_offs should be considered.
+    cl_ulong offset;
+    // The actual size of the cl_mem object. This is needed when returning the
+    // block to the pool.
+    size_t actual_size;
+
+    void reset() {
+        data_device = nullptr;
+        offset = 0;
+        actual_size = 0;
+    }
+};
+
+// Additional tensor extra structs for quantized tensors.
+// These tensors are loaded from files and should not be allocated in scratch --
+// they should always be allocated from the pool. Hence, they do not have an
+// `offset`, which indicate their locations in the scratch buffer.
+struct ggml_tensor_extra_cl_q4_0 {
+    // Quantized values.
+    cl_mem q = nullptr;
+    // Quantized values in image1d_buffer_t.
+    cl_mem q_img = nullptr;
+    // Scales.
+    cl_mem d = nullptr;
+    // Scales in image1d_buffer_t.
+    cl_mem d_img = nullptr;
+    // Size of quantized values.
+    size_t size_q = 0;
+    // Size of scales.
+    size_t size_d = 0;
+
+    ~ggml_tensor_extra_cl_q4_0() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        // Currently, q_img and d_img are only initialized when SMALL_ALLOC is
+        // enabled. They point to the images in ggml_backend_opencl_buffer_context.
+        // So, there is no need to release them here.
+        // TODO: initialize them for non SMALL_PATH path, or remove them.
+        q_img = nullptr;
+        d_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+    }
+};
+
+struct ggml_tensor_extra_cl_mxfp4 {
+    // Quantized values.
+    cl_mem q = nullptr;
+    // Quantized values in image1d_buffer_t.
+    cl_mem q_img = nullptr;
+    // Scales in E8M0.
+    cl_mem e = nullptr;
+    // Scales in image1d_buffer_t.
+    cl_mem e_img = nullptr;
+    // Size of quantized values.
+    size_t size_q = 0;
+    // Size of scales.
+    size_t size_e = 0;
+
+    ~ggml_tensor_extra_cl_mxfp4() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (e != nullptr) {
+            CL_CHECK(clReleaseMemObject(e));
+            e = nullptr;
+        }
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q_img));
+            q = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        e_img = nullptr;
+        size_q = 0;
+        size_e = 0;
+    }
+};
+
+struct ggml_tensor_extra_cl_q8_0 {
+    cl_mem q = nullptr;
+    cl_mem q_img = nullptr;
+
+    cl_mem d = nullptr;
+    cl_mem d_img = nullptr;
+
+    size_t size_q = 0;
+    size_t size_d = 0;
+
+    ~ggml_tensor_extra_cl_q8_0() {
+        reset();
+    }
+
+    void reset() {
+        // q and d are subbuffers into the bigger buffer allocated in ggml_backend_buffer.
+        // They must be properly released so that the original buffer can be
+        // properly released to avoid memory leak.
+        if (q != nullptr) {
+            CL_CHECK(clReleaseMemObject(q));
+            q = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+        // Currently, q_img and d_img are not used. They can be image1d_buffer_t
+        // that wraps around q and d to utilize image access path.
+        q_img = nullptr;
+        d_img = nullptr;
+        size_q = 0;
+        size_d = 0;
+    }
+};
+
+struct ggml_tensor_extra_cl_q6_K {
+    // Lower 4 bits of quantized weights.
+    cl_mem ql = nullptr;
+    // Upper 2 bits of quantized weights.
+    cl_mem qh = nullptr;
+    // Scales for each block.
+    cl_mem s  = nullptr;
+    // Scales for each super block.
+    cl_mem d  = nullptr;
+
+    size_t size_ql = 0;
+    size_t size_qh = 0;
+    size_t size_s  = 0;
+    size_t size_d  = 0;
+
+    ~ggml_tensor_extra_cl_q6_K() {
+        reset();
+    }
+
+    void reset() {
+        if (ql != nullptr) {
+            CL_CHECK(clReleaseMemObject(ql));
+            ql = nullptr;
+        }
+        if (qh != nullptr) {
+            CL_CHECK(clReleaseMemObject(qh));
+            qh = nullptr;
+        }
+        if (s != nullptr) {
+            CL_CHECK(clReleaseMemObject(s));
+            s = nullptr;
+        }
+        if (d != nullptr) {
+            CL_CHECK(clReleaseMemObject(d));
+            d = nullptr;
+        }
+
+        size_ql = 0;
+        size_qh = 0;
+        size_s  = 0;
+        size_d  = 0;
+    }
+};
+
+//------------------------------------------------------------------------------
+// Backend API
+//------------------------------------------------------------------------------
+
+//
+// backend
+//
+static const char * ggml_backend_opencl_name(ggml_backend_t backend) {
+    return "OpenCL";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_opencl_free(ggml_backend_t backend) {
+    ggml_cl2_free(backend);
+}
+
+static void ggml_backend_opencl_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_UNUSED(backend);
+    GGML_UNUSED(tensor);
+    GGML_UNUSED(data);
+    GGML_UNUSED(offset);
+    GGML_UNUSED(size);
+}
+
+static void ggml_backend_opencl_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_UNUSED(backend);
+    GGML_UNUSED(tensor);
+    GGML_UNUSED(data);
+    GGML_UNUSED(offset);
+    GGML_UNUSED(size);
+}
+
+static bool ggml_backend_opencl_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src);
+    GGML_UNUSED(dst);
+    return false;
+}
+
+static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+
+    cl_event evt;
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, 0, nullptr, &evt));
+    CL_CHECK(clWaitForEvents(1, &evt));
+    CL_CHECK(clReleaseEvent(evt));
+}
+
+// Syncronizes the 'backend_ctx's device with others so that commands
+// enqueued to it won't start until commands in the other devices have
+// completed.
+static void sync_with_other_backends(ggml_backend_opencl_context * backend_ctx) {
+    if (g_ggml_backend_opencl_devices.size() < 2)
+      return; // No other devices to synchronize with.
+
+    std::vector<cl_event> events;
+    events.reserve(g_ggml_backend_opencl_devices.size());
+
+    for (ggml_backend_device & backend_dev : g_ggml_backend_opencl_devices) {
+        auto * other_backend_ctx = ggml_cl2_init(&backend_dev);
+        if (backend_ctx != other_backend_ctx) {
+            cl_event ev;
+            CL_CHECK(clEnqueueMarkerWithWaitList(other_backend_ctx->queue, 0, nullptr, &ev));
+            CL_CHECK(clFlush(other_backend_ctx->queue));
+            events.push_back(ev);
+        }
+    }
+
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, events.size(), events.data(), nullptr));
+    for (auto ev : events) {
+        CL_CHECK(clReleaseEvent(ev));
+    }
+}
+
+static void sync_with_other_backends(ggml_backend_t backend) {
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+    sync_with_other_backends(backend_ctx);
+}
+
+static bool ggml_opencl_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops) {
+    if (!ggml_can_fuse(cgraph, node_idx, ops)) {
+        return false;
+    }
+
+    if (ops.size() == 2 && ops.begin()[0] == GGML_OP_RMS_NORM && ops.begin()[1] == GGML_OP_MUL) {
+        const ggml_tensor *rms_norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul      = cgraph->nodes[node_idx+1];
+
+        GGML_ASSERT(rms_norm->src[0]->type == GGML_TYPE_F32);
+        GGML_ASSERT(rms_norm->type == GGML_TYPE_F32);
+
+        // rms_norm only supports f32
+        if (mul->src[0]->type != GGML_TYPE_F32 ||
+            mul->src[1]->type != GGML_TYPE_F32 ||
+            mul->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        // if rms_norm is the B operand, then we don't handle broadcast
+        if (rms_norm == mul->src[1] &&
+            !ggml_are_same_shape(mul->src[0], rms_norm)) {
+            return false;
+        }
+
+        // rms_norm assumes contiguous rows
+        if (!ggml_is_contiguous_rows(mul->src[0]) || !ggml_is_contiguous_rows(mul->src[1])) {
+            return false;
+        }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *norm = cgraph->nodes[node_idx];
+        const ggml_tensor *mul  = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add  = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w    = mul->src[0] == norm ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b    = add->src[0] == mul  ? add->src[1] : add->src[0];
+
+        // norm fusion only supports F32
+        if (norm->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        if (norm->src[0]->ne[0] % 4 != 0) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(norm->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
+    } else if (ops.size() == 3 && ops.begin()[0] == GGML_OP_GROUP_NORM && ops.begin()[1] == GGML_OP_MUL && ops.begin()[2] == GGML_OP_ADD) {
+        const ggml_tensor *gn = cgraph->nodes[node_idx];
+        const ggml_tensor *mul = cgraph->nodes[node_idx+1];
+        const ggml_tensor *add = cgraph->nodes[node_idx+2];
+        const ggml_tensor *w   = mul->src[0] == gn ? mul->src[1] : mul->src[0];
+        const ggml_tensor *b   = add->src[0] == mul ? add->src[1] : add->src[0];
+
+        if (gn->src[0]->type != GGML_TYPE_F32 || w->type != GGML_TYPE_F32 || b->type != GGML_TYPE_F32) {
+            return false;
+        }
+
+        if (!ggml_is_contiguous(gn->src[0]) || !ggml_is_contiguous(w) || !ggml_is_contiguous(b)) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor * rms_norm_tensor, ggml_tensor * mul_tensor);
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor);
+
+static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor * node = cgraph->nodes[i];
+
+        // NOTE: this may oversynchronize by synchronizing with
+        //       backends/devices which don't compute 'cgraph's
+        //       dependencies.
+        sync_with_other_backends(backend);
+
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+            continue;
+        }
+
+        if ((node->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+            continue;
+        }
+
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_GROUP_NORM, GGML_OP_MUL, GGML_OP_ADD })) {
+            ggml_opencl_op_group_norm_fused(backend, node, cgraph->nodes[i+1], cgraph->nodes[i+2]);
+            i += 2;
+            continue;
+        }
+        if (!backend_ctx->disable_fusion && ggml_opencl_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+            ggml_opencl_op_rms_norm_fused(backend, node, cgraph->nodes[i+1]);
+            i++;
+            continue;
+        }
+
+        bool ok = ggml_cl_compute_forward(backend, node);
+        if (!ok) {
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    ggml_backend_opencl_device_context * dev_ctx     = (ggml_backend_opencl_device_context *)dev->context;
+    ggml_backend_opencl_context *        backend_ctx = dev_ctx->backend_ctx;
+
+    switch (op->op) {
+        case GGML_OP_NONE:
+            return true;
+        case GGML_OP_GET_ROWS:
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    return true;
+                case GGML_TYPE_Q4_0:
+#ifdef GGML_OPENCL_SOA_Q
+                    // We do not support flattened Q4_0 (and possibly other Q's)
+                    return false;
+#else // GGML_OPENCL_SOA_Q
+                    return true;
+#endif // GGML_OPENCL_SOA_Q
+                default:
+                    return false;
+            }
+        case GGML_OP_SET_ROWS:
+            {
+                // TODO: add support
+                // ref: https://github.com/ggml-org/llama.cpp/pull/14274
+#pragma message("TODO: implement BF16, Q4_0, Q4_1, Q5_0, Q5_1, Q8_0, IQ4_NL support (https://github.com/ggml-org/llama.cpp/pull/14661)")
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+                switch (op->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                        return (op->src[1]->type == GGML_TYPE_I64 || op->src[1]->type == GGML_TYPE_I32);
+                    default:
+                        return false;
+                }
+            }
+        case GGML_OP_CPY:
+        case GGML_OP_DUP:
+        case GGML_OP_CONT:
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                    switch (op->type) {
+                        case GGML_TYPE_F16:
+                        case GGML_TYPE_F32:
+                            return true;
+                        default:
+                            return false;
+                    }
+                case GGML_TYPE_F16:
+                    switch (op->type) {
+                        case GGML_TYPE_F16:
+                        case GGML_TYPE_F32:
+                            return true;
+                        default:
+                            return false;
+                    }
+                default:
+                    return false;
+            }
+        case GGML_OP_SCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_ADD:
+            if (op->type == GGML_TYPE_F16) {
+                const bool src0_ok = op->src[0]->type == GGML_TYPE_F16 || op->src[0]->type == GGML_TYPE_F32;
+                const bool src1_ok = op->src[1]->type == GGML_TYPE_F16 || op->src[1]->type == GGML_TYPE_F32;
+                if (src0_ok && src1_ok) {
+                    return true;
+                }
+            }
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SUB:
+            return (op->src[0]->type == op->src[1]->type) &&
+                   (op->src[0]->type == op->type) &&
+                   (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16);
+        case GGML_OP_ADD_ID:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+            return (op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16) &&
+                    ggml_is_contiguous(op->src[0]);
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU_ERF:
+                case GGML_UNARY_OP_GELU_QUICK:
+                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                case GGML_UNARY_OP_SIGMOID:
+                    return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_TANH:
+                   return op->src[0]->type == GGML_TYPE_F32 || op->src[0]->type == GGML_TYPE_F16;
+                case GGML_UNARY_OP_EXPM1:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                case GGML_UNARY_OP_SOFTPLUS:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                default:
+                    return false;
+            }
+        case GGML_OP_GLU:
+            switch (ggml_get_glu_op(op)) {
+                case GGML_GLU_OP_GEGLU:
+                case GGML_GLU_OP_REGLU:
+                case GGML_GLU_OP_SWIGLU:
+                case GGML_GLU_OP_SWIGLU_OAI:
+                case GGML_GLU_OP_GEGLU_ERF:
+                case GGML_GLU_OP_GEGLU_QUICK:
+                    return ggml_is_contiguous_1(op->src[0]) && (op->type == GGML_TYPE_F32 || op->type == GGML_TYPE_F16);
+                default:
+                    return false;
+            }
+        case GGML_OP_TRI:
+            return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
+        case GGML_OP_FILL:
+            return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
+        case GGML_OP_CLAMP:
+            return op->src[0]->type == GGML_TYPE_F32;
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_NORM:
+            return true;
+        case GGML_OP_RMS_NORM:
+            return op->ne[0] % 4 == 0 && ggml_is_contiguous_rows(op->src[0]);
+        case GGML_OP_REPEAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
+        case GGML_OP_PAD:
+            // TODO: add circular padding support for opencl, see https://github.com/ggml-org/llama.cpp/pull/16985
+            if (ggml_get_op_params_i32(op, 8) != 0) {
+                return false;
+            }
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_UPSCALE: {
+            ggml_scale_mode mode = (ggml_scale_mode)(ggml_get_op_params_i32(op, 0) & 0xFF);
+            const bool antialias = (ggml_scale_mode)(ggml_get_op_params_i32(op, 0) & GGML_SCALE_FLAG_ANTIALIAS);
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   (mode == GGML_SCALE_MODE_NEAREST || mode == GGML_SCALE_MODE_BILINEAR) && !antialias;
+        }
+        case GGML_OP_CONV_2D:
+            return (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16) ||
+                   (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                   (op->src[0]->type == GGML_TYPE_F16 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_SSM_CONV:
+            return (op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32);
+        case GGML_OP_CONCAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_GROUP_NORM:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_MUL_MAT:
+            if (op->src[0]->type == GGML_TYPE_F16) {
+                return true;
+            } else if (op->src[0]->type == GGML_TYPE_F32) {
+                return op->src[1]->type == GGML_TYPE_F32;
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
+                       op->src[0]->type == GGML_TYPE_Q4_K ||
+                       op->src[0]->type == GGML_TYPE_Q6_K) {
+                return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+            } else if (op->src[0]->type == GGML_TYPE_Q8_0) {
+                return op->src[1]->type == GGML_TYPE_F32;
+            }
+            return false;
+        case GGML_OP_MUL_MAT_ID:
+            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_Q8_0 ||
+                op->src[0]->type == GGML_TYPE_MXFP4) {
+                if (op->src[1]->type == GGML_TYPE_F32) {
+                    return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
+                }
+            }
+            return false;
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            return true;
+        case GGML_OP_DIAG_MASK_INF:
+            return op->ne[3] == 1;
+        case GGML_OP_ROPE: {
+            const int mode = ((const int32_t *) op->op_params)[2];
+            const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+            const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+            if (is_mrope && !is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
+                return false;
+            }
+            if (is_vision) {
+                if (op->src[0]->type == GGML_TYPE_F32 ||
+                    op->src[0]->type == GGML_TYPE_F16) {
+                    return true;
+                }
+                return false;
+            }
+            return true;
+        }
+        case GGML_OP_SOLVE_TRI:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_IM2COL:
+            return true;
+        case GGML_OP_ARGSORT: {
+            cl_kernel kernel = backend_ctx->kernel_argsort_f32_i32;
+            int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+
+            int cols = 1;
+            while (cols < op->ne[0]) {
+                cols *= 2;
+            }
+
+            return cols <= max_workgroup_size && op->src[0]->type == GGML_TYPE_F32;
+        }
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
+            return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                const ggml_tensor * q = op->src[0];
+                const ggml_tensor * k = op->src[1];
+                const ggml_tensor * v = op->src[2];
+
+                const int dk = q->ne[0];
+                const int dv = v->ne[0];
+
+                const struct { int dk; int dv; } supported_dims[] = {
+                    { 40,  40}, { 64,  64}, { 80,  80}, { 96,  96},
+                    {112, 112}, {128, 128}, {192, 128},
+                    {192, 192}, {256, 256},
+                };
+
+                bool dims_supported = false;
+                for (size_t i = 0; i < sizeof(supported_dims)/sizeof(supported_dims[0]); ++i) {
+                    if (supported_dims[i].dk == dk && supported_dims[i].dv == dv) {
+                        dims_supported = true;
+                        break;
+                    }
+                }
+                if (!dims_supported) {
+                    return false;
+                }
+
+                const bool is_f32_f32 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F32 &&
+                                        v->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+                const bool is_f16_f16 = q->type == GGML_TYPE_F16 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
+                const bool is_f32_f16 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
+                return is_f32_f32 || is_f16_f16 || is_f32_f16;
+            }
+        default:
+            return false;
+    }
+}
+
+// Forward declaration - implementation appears later in the file.
+static const char * ggml_backend_opencl_buffer_type_get_name(ggml_backend_buffer_type_t buffer_type);
+
+static ggml_guid_t ggml_backend_opencl_guid() {
+    static ggml_guid guid = { 0xde, 0xe0, 0x70, 0xa2, 0x73, 0x4e, 0x4d, 0xbc, 0xb0, 0xc7, 0x4f, 0xd4, 0x6d, 0x4e, 0x90, 0xfe };
+    return &guid;
+}
+
+static ggml_backend_i ggml_backend_opencl_i = {
+    /* .get_name                = */ ggml_backend_opencl_name,
+    /* .free                    = */ ggml_backend_opencl_free,
+    /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
+    /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
+    /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
+    /* .synchronize             = */ ggml_backend_opencl_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_opencl_graph_compute,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .graph_optimize          = */ NULL,
+};
+
+ggml_backend_t ggml_backend_opencl_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_opencl_reg(), 0);
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid    = */ ggml_backend_opencl_guid(),
+        /* .iface   = */ ggml_backend_opencl_i,
+        /* .device  = */ dev,
+        /* .context = */ backend_ctx
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_opencl(ggml_backend_t backend) {
+    return backend && backend->iface.get_name == ggml_backend_opencl_name;
+}
+
+//
+// buffer
+//
+struct ggml_backend_opencl_buffer_context {
+    // A buffer context can hold multiple cl_mem objects. This is for flattening
+    // quantized weights and should be used with GGML_OPENCL_SMALL_ALLOC where
+    // each tensor is allocated a separate buffer. When flattening is enabled
+    // with small allocation, each tensor is backed by two cl_mem objects (for
+    // quants and scales) packed into a backend_opencl_buffer.
+    ggml_backend_opencl_buffer_context(cl_mem buf)
+        : name("OpenCL") {
+        buffer.push_back(buf);
+    }
+
+    ~ggml_backend_opencl_buffer_context() {
+        for (cl_mem buf : buffer) {
+            CL_CHECK(clReleaseMemObject(buf));
+        }
+        for (cl_mem im : img) {
+            CL_CHECK(clReleaseMemObject(im));
+        }
+
+        // Delete all extras to trigger their destructors
+        for (ggml_tensor_extra_cl * e : temp_tensor_extras) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q4_0 * e : temp_tensor_extras_q4_0) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q4_0 * e : temp_tensor_extras_q4_0_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K) {
+            delete e;
+        }
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            delete e;
+        }
+    }
+
+    ggml_tensor_extra_cl * ggml_opencl_alloc_temp_tensor_extra() {
+        ggml_tensor_extra_cl * extra;
+        if (temp_tensor_extras.empty()) {
+            extra = new ggml_tensor_extra_cl();
+        } else {
+            extra = temp_tensor_extras.back();
+            temp_tensor_extras.pop_back();
+        }
+
+        temp_tensor_extras_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_q4_0 * ggml_opencl_alloc_temp_tensor_extra_q4_0() {
+        ggml_tensor_extra_cl_q4_0 * extra;
+        if (temp_tensor_extras_q4_0.empty()) {
+            extra = new ggml_tensor_extra_cl_q4_0();
+        } else {
+            extra = temp_tensor_extras_q4_0.back();
+            temp_tensor_extras_q4_0.pop_back();
+        }
+
+        temp_tensor_extras_q4_0_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_mxfp4 * ggml_opencl_alloc_temp_tensor_extra_mxfp4() {
+        ggml_tensor_extra_cl_mxfp4 * extra;
+        if (temp_tensor_extras_mxfp4.empty()) {
+            extra = new ggml_tensor_extra_cl_mxfp4();
+        } else {
+            extra = temp_tensor_extras_mxfp4.back();
+            temp_tensor_extras_mxfp4.pop_back();
+        }
+
+        temp_tensor_extras_mxfp4_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_q8_0 * ggml_opencl_alloc_temp_tensor_extra_q8_0() {
+        ggml_tensor_extra_cl_q8_0 * extra;
+        if (temp_tensor_extras_q8_0.empty()) {
+            extra = new ggml_tensor_extra_cl_q8_0();
+        } else {
+            extra = temp_tensor_extras_q8_0.back();
+            temp_tensor_extras_q8_0.pop_back();
+        }
+
+        temp_tensor_extras_q8_0_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    ggml_tensor_extra_cl_q6_K * ggml_opencl_alloc_temp_tensor_extra_q6_K() {
+        ggml_tensor_extra_cl_q6_K * extra;
+        if (temp_tensor_extras_q6_K.empty()) {
+            extra = new ggml_tensor_extra_cl_q6_K();
+        } else {
+            extra = temp_tensor_extras_q6_K.back();
+            temp_tensor_extras_q6_K.pop_back();
+        }
+
+        temp_tensor_extras_q6_K_in_use.push_back(extra);
+
+        extra->reset();
+        return extra;
+    }
+
+    void reset() {
+        for (ggml_tensor_extra_cl * e : temp_tensor_extras_in_use) {
+            temp_tensor_extras.push_back(e);
+        }
+        temp_tensor_extras_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q4_0 * e : temp_tensor_extras_q4_0_in_use) {
+            temp_tensor_extras_q4_0.push_back(e);
+        }
+        temp_tensor_extras_q4_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_mxfp4 * e : temp_tensor_extras_mxfp4_in_use) {
+            temp_tensor_extras_mxfp4.push_back(e);
+        }
+        temp_tensor_extras_mxfp4_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q8_0 * e : temp_tensor_extras_q8_0_in_use) {
+            temp_tensor_extras_q8_0.push_back(e);
+        }
+        temp_tensor_extras_q8_0_in_use.clear();
+
+        for (ggml_tensor_extra_cl_q6_K * e : temp_tensor_extras_q6_K_in_use) {
+            temp_tensor_extras_q6_K.push_back(e);
+        }
+        temp_tensor_extras_q6_K_in_use.clear();
+    }
+
+    // Pools for extras. Available extras are in `temp_tensor_extras`. Extras
+    // being used are in `temp_tensor_extras_in_use`. At the first run, new
+    // extras get created and put in `in_use`. When the buffer is reset via
+    // the `reset` callback, all extras in `in_use` get moved to available extras
+    // for reuse.
+    std::vector<ggml_tensor_extra_cl *> temp_tensor_extras;
+    std::vector<ggml_tensor_extra_cl *> temp_tensor_extras_in_use;
+    std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0;
+    std::vector<ggml_tensor_extra_cl_q4_0 *> temp_tensor_extras_q4_0_in_use;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4;
+    std::vector<ggml_tensor_extra_cl_mxfp4 *> temp_tensor_extras_mxfp4_in_use;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0;
+    std::vector<ggml_tensor_extra_cl_q8_0 *> temp_tensor_extras_q8_0_in_use;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K;
+    std::vector<ggml_tensor_extra_cl_q6_K *> temp_tensor_extras_q6_K_in_use;
+
+    // The buffer_context is initially created by ggml_backend_buft_alloc_buffer
+    // before any tensor is initialized (at the beginning of alloc_tensor_range).
+    // Hence, there is alway a buffer object in this vector. When each tensor is
+    // being initialized, this original buffer object will be released if both
+    // flattening and small allocation are enabled, and additional buffer
+    // objects will be created in init_tensor to represent flattened quantized
+    // weights.
+    std::vector<cl_mem> buffer;
+    // These are image1d_buffer_t objects that wrap around the quants and scales.
+    // For Q4_0 quantization, there should be two of them - one for quants and
+    // one for scales. They should be populated only when flattening and small
+    // allocation are enabled.
+    std::vector<cl_mem> img;
+    std::string name;
+};
+
+static void ggml_backend_opencl_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_opencl_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer->buft->device);
+    return (void *) (uintptr_t) backend_ctx->alignment;
+}
+
+static enum ggml_status ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+
+    ggml_cl2_init(buffer->buft->device);
+
+    if (tensor->view_src != nullptr) {
+        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
+
+        ggml_tensor_extra_cl * view_extra = (ggml_tensor_extra_cl *) tensor->view_src->extra;
+        GGML_ASSERT(view_extra && "view_extra is nullptr?");
+
+        // Reuse extra of the parent tensor. The offset of this view tensor
+        // becomes `extra->offset + view_offs` and needs to be calculated when
+        // it is used. This changes is needed because of the change to
+        // ggml_alloc.c in https://github.com/ggml-org/llama.cpp/pull/7640.
+        // `buffer` passed in here will always be `tensor->buffer`. It is OK
+        // to allocate extras from the same buffer context for ordinary
+        // intermediate tensors. But for views into kv cache tensors, doing so
+        // would mess up the extras used by kv cache.
+        // Before #7640, `buffer` is for intermediate tensors, which is always
+        // different from that of kv cache tensors.
+        //
+        // NB: now extra->offset no longer accounts for view_offs.
+        // NB: this should not apply to weight tensors (for end-to-end runs, but
+        //     may apply for test-backend-ops).
+        // FIXME: if any unexpected results are seen, double check the offset -
+        // there could be other places that need fix.
+        tensor->extra = view_extra;
+    } else {
+        {
+            size_t offset = (char *) tensor->data - (char *) ggml_backend_opencl_buffer_get_base(buffer);
+
+            ggml_tensor_extra_cl * extra = ctx->ggml_opencl_alloc_temp_tensor_extra();
+            extra->offset = offset;
+            extra->data_device = ctx->buffer[0];
+            extra->actual_size = ggml_nbytes(tensor);
+
+            tensor->extra = extra;
+        }
+    }
+    return GGML_STATUS_SUCCESS;
+}
+
+// The optimized gemm and gemv kernels are used for large matrices without batch.
+// tensor is the quantized weights matrix.
+inline bool use_adreno_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+    int64_t threshold_ne0 = 512;
+    int64_t threshold_ne1 = 512;
+    if (!backend_ctx->adreno_cl_compiler_version.newer_than_or_same(E031, 38, 11, 0) &&
+         backend_ctx->adreno_cl_compiler_version.type != DX) {
+        threshold_ne0 = 128;
+        threshold_ne1 = 128;
+    }
+    return tensor->ne[0] >= threshold_ne0 && tensor->ne[1] >= threshold_ne1 &&
+            tensor->ne[2] == 1 && tensor->ne[3] == 1;
+}
+
+inline bool use_adreno_moe_kernels(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+    GGML_UNUSED(backend_ctx);
+    int ne01 = tensor->ne[1];
+    return ((strstr(tensor->name, "ffn") != NULL) || (strstr(tensor->name, "as") != NULL)) && (ne01 % 64 == 0);
+}
+
+inline bool enable_adreno_trans_weight(const ggml_backend_opencl_context *backend_ctx, const ggml_tensor *tensor) {
+
+    bool adreno_kernel = use_adreno_kernels(backend_ctx, tensor);
+
+    size_t elem_num = tensor->ne[0] * tensor->ne[1] * tensor->ne[2] * tensor->ne[3];
+
+    return ((elem_num < 128 * 1024 * 1024) && adreno_kernel);  // max element num: 2**27
+}
+
+static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(buffer->buft->device);
+
+    cl_context context = backend_ctx->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+#ifdef GGML_OPENCL_SOA_Q
+    // We separate the quantized bits and scale from block_q4_0 by using an
+    // additional kernel, where each thread handles a block. We first read the
+    // original weights into a temporary buffer, then create two separate
+    // buffers for quantized bits and scales, which are then populated by the
+    // conversion kernel.
+    if (tensor->type == GGML_TYPE_Q4_0) {
+        // Tensors should have been preallocated, therefore they should
+        // already have ggml_tensor_extra_cl as extra.
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q4_0 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q4_0();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // We consider the specified offset arg as always, although For weights
+        // the offset arg should be 0 (we do not assert this).
+        //GGML_ASSERT(offset == 0);
+
+        // We create subbuffers from the original tensor buffer for scales and
+        // quants - i.e., scales and quants are aliases into the buffer obejct
+        // that backs the original tensor. This is a cleaner way to adapt to the
+        // new memory management.
+        // In the old code, we allocate new buffers for scales and quants
+        // respectively, which could still be done but would result in double
+        // allocation; properly deallocating the preallocated buffer that backs
+        // the tensors is tricky and would leak the backend specific information
+        // into the general backend code.
+        // Does this create misaligned subbuffers (alignment is 1024) in certain
+        // cases ?
+        cl_buffer_region region;
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        //cl_kernel kernel = backend_ctx->kernel_convert_block_q4_0;
+    #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_0;
+
+        // The optimized kernels need weights in natural order, so unshuffle.
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_q4_0_noshuffle;
+        }
+    #else
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q4_0;
+    #endif // GGML_OPENCL_USE_ADRENO_KERNELS
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
+        // transpose the weights and scales
+    #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        // Only do transpose for large, non batched matrix
+        // TODO: use preallocated images instead of sub-buffer then image
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+        // <----------------------------------------------------------------------------------> //
+        // start transpose
+        // <----------------------------------------------------------------------------------> //
+        int M = tensor->ne[1];   // ne01
+        int K = tensor->ne[0];   // ne00
+
+        //For matrix-vector multiplication kernel, we assume K is a multiple of 32
+        GGML_ASSERT(K % 32 == 0);
+        //For transpose kernels, we assume K is a multiple of 4 (satisfied by prior assert), and M is a multiple of 4
+        GGML_ASSERT(M % 4 == 0);
+
+        // transpose is out of place, so we need to allocate transposed buffers
+        // <----------------------------------------------------------------------------------> //
+        // use sub_buffer of max buffer size instead
+
+        size_t q_size_bytes = K * M / 8 * sizeof(float);
+        backend_ctx->prealloc_quant_trans.allocate(context, q_size_bytes);
+
+        cl_buffer_region region;
+        region.origin = 0;
+        region.size = q_size_bytes;
+        cl_mem qT_d = clCreateSubBuffer(
+            backend_ctx->prealloc_quant_trans.buffer,
+            0,
+            CL_BUFFER_CREATE_TYPE_REGION,
+            &region,
+            &err);
+        CL_CHECK(err);
+
+        bool K_tile_trans = true;
+        if ((K / 32) % 4 != 0){
+            K_tile_trans =false;
+        }
+
+        size_t d_size_bytes = M * (K / 32) * 2;
+        backend_ctx->prealloc_scales_trans.allocate(context, d_size_bytes);
+
+        region.origin = 0;
+        region.size = d_size_bytes;
+        cl_mem dT_d = clCreateSubBuffer(
+            backend_ctx->prealloc_scales_trans.buffer,
+            0,
+            CL_BUFFER_CREATE_TYPE_REGION,
+            &region,
+            &err);
+        CL_CHECK(err);
+
+        // <----------------------------------------------------------------------------------> //
+
+
+        // create images from the buffers
+        // <----------------------------------------------------------------------------------> //
+        cl_mem q_d_image1D;
+        cl_mem d_d_image1D;
+        cl_mem qT_d_image1D;
+        cl_mem dT_d_image1D;
+
+        cl_image_format img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+        cl_image_desc img_desc_1d;
+
+        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_1d.image_width = M * K / 4 / 4;
+        img_desc_1d.buffer = extra->q;
+        q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+        CL_CHECK(err);
+
+        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_1d.image_width = M * K / 4 / 4;
+        img_desc_1d.buffer = qT_d;
+        qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+        CL_CHECK(err);
+
+        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        if (K_tile_trans) {
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32 / 4;
+        } else {
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+        }
+        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_1d.buffer = extra->d;
+        d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+        CL_CHECK(err);
+
+        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_1d.image_width = M * K / 32 / 4;
+        img_desc_1d.buffer = dT_d;
+        dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+        CL_CHECK(err);
+        // <----------------------------------------------------------------------------------> //
+
+        // set up and call the transpose kernels
+        // <----------------------------------------------------------------------------------> //
+        // weights
+        int height_q = M / 4;
+        int width_q = K / 4 / 4;
+        kernel = backend_ctx->kernel_transpose_16;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &qT_d_image1D));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_q));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_q));
+
+        size_t local_size_q[3] = {4, 16, 1};
+        size_t global_size_q[3] = {static_cast<size_t>(width_q), static_cast<size_t>(height_q), 1};
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_q, local_size_q, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+
+        // scales
+        int height_s = M / 4;
+        int width_s = K / 32 / 4;
+
+        kernel = backend_ctx->kernel_transpose_16;
+        if (!K_tile_trans) {
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+            width_s = K / 32;
+        }
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &width_s));
+
+        size_t local_size_s[3] = {4, 16, 1};
+        size_t global_size_s[3] = {static_cast<size_t>(width_s), static_cast<size_t>(height_s), 1};
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_s, local_size_s, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        // <----------------------------------------------------------------------------------> //
+
+        // copy transposed buffer contents to original buffers
+        // <----------------------------------------------------------------------------------> //
+        // weights
+        CL_CHECK(clEnqueueCopyBuffer(queue, qT_d, extra->q, 0, 0, q_size_bytes, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+
+        // scales
+        CL_CHECK(clEnqueueCopyBuffer(queue, dT_d, extra->d, 0, 0, d_size_bytes, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        // <----------------------------------------------------------------------------------> //
+
+        // deallocate transpose buffers
+        // <----------------------------------------------------------------------------------> //
+        CL_CHECK(clReleaseMemObject(qT_d));
+        CL_CHECK(clReleaseMemObject(dT_d));
+
+        // deallocate temporary images
+        CL_CHECK(clReleaseMemObject(q_d_image1D));
+        CL_CHECK(clReleaseMemObject(d_d_image1D));
+        CL_CHECK(clReleaseMemObject(qT_d_image1D));
+        CL_CHECK(clReleaseMemObject(dT_d_image1D));
+        // <----------------------------------------------------------------------------------> //
+        // end transpose
+        // <----------------------------------------------------------------------------------> //
+        }
+    #endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+        return;
+
+    }
+    if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_mxfp4 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_mxfp4();
+
+        size_t size_e = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(char);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        GGML_ASSERT(size_e + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_e;
+        extra->e = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_e, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_moe_kernels(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4_trans;
+
+            int ne00 = tensor->ne[0];
+            int ne01 = tensor->ne[1];
+            int ne02 = tensor->ne[2];
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &ne00));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int), &ne01));
+
+            size_t global_work_size[3] = {static_cast<size_t>(((ne01 + 63) / 64) * 64), static_cast<size_t>(ne00 / 32), static_cast<size_t>(ne02)};
+            size_t local_work_size[3] = {64, 2, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+            CL_CHECK(clReleaseMemObject(data_device));
+            tensor->extra = extra;
+
+            return;
+        }
+#endif
+        cl_kernel kernel = backend_ctx->kernel_convert_block_mxfp4;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->e));
+
+        size_t global_work_size[3] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[3] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        // Create image for Q
+        cl_image_format img_format_q = {CL_RG, CL_UNSIGNED_INT32};
+        cl_image_desc img_desc_q = {
+            CL_MEM_OBJECT_IMAGE1D_BUFFER,
+            static_cast<size_t>(ggml_nelements(tensor)/32*2),
+            0, 0, 0, 0, 0, 0, 0,
+            { extra->q }
+        };
+        extra->q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_format_q, &img_desc_q, NULL, &err);
+        tensor->extra = extra;
+
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q8_0 * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q8_0();
+
+        size_t size_d = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        size_t size_q = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*(ggml_blck_size(tensor->type)*sizeof(char));
+        GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        // The original tensor memory is divided into scales and quants, i.e.,
+        // we first store scales, then quants.
+        cl_buffer_region region;
+
+        // Create subbuffer for scales.
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Create subbuffer for quants.
+        region.origin = align_to(previous_origin + size_d, backend_ctx->alignment);
+        region.size = size_q;
+        extra->q = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q8_0;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        tensor->extra = extra;
+
+        // Transpose the weights and scales
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (enable_adreno_trans_weight(backend_ctx, tensor)) {
+
+            int M = tensor->ne[1];   // ne01
+            int K = tensor->ne[0];   // ne00
+
+            GGML_ASSERT(K % 32 == 0);
+            GGML_ASSERT(M % 4 == 0);
+            GGML_ASSERT(tensor->ne[2] == 1);
+            GGML_ASSERT(tensor->ne[3] == 1);
+
+            // Transpose weights
+            size_t q_size_bytes = K * M / 4 * sizeof(float);
+            cl_buffer_region region;
+            region.origin = 0;
+            region.size = q_size_bytes;
+            cl_mem qT_d = clCreateSubBuffer(
+                backend_ctx->prealloc_quant_trans.buffer,
+                0,
+                CL_BUFFER_CREATE_TYPE_REGION,
+                &region,
+                &err);
+            CL_CHECK(err);
+
+            cl_mem q_d_image1D;
+            cl_mem qT_d_image1D;
+
+            cl_image_format img_fmt_1d;
+            cl_image_desc img_desc_1d;
+
+            img_fmt_1d = { CL_RGBA, CL_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 4 / 4;
+            img_desc_1d.buffer = extra->q;
+            q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            img_fmt_1d = { CL_RGBA, CL_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 4 / 4;
+            img_desc_1d.buffer = qT_d;
+            qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            int height_q = M / 4;
+            int width_q = K / 4 / 4;
+            kernel = backend_ctx->kernel_transpose_32;
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &qT_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_q));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_q));
+
+            size_t local_size_q[3] = {4, 16, 1};
+            size_t global_size_q[3] = {static_cast<size_t>(width_q), static_cast<size_t>(height_q), 1};
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_q, local_size_q, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            // Transpose scales
+            size_t d_size_bytes = M * (K / 32) * 2;
+            region.origin = 0;
+            region.size = d_size_bytes;
+            cl_mem dT_d = clCreateSubBuffer(
+                backend_ctx->prealloc_scales_trans.buffer,
+                0,
+                CL_BUFFER_CREATE_TYPE_REGION,
+                &region,
+                &err);
+            CL_CHECK(err);
+
+            cl_mem d_d_image1D;
+            cl_mem dT_d_image1D;
+
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.buffer = extra->d;
+            d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.image_width = M * K / 32 / 4;
+            img_desc_1d.buffer = dT_d;
+            dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
+            CL_CHECK(err);
+
+            int height_s = M / 4;
+            int width_s = K / 32;
+
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &width_s));
+
+            size_t local_size_s[3] = {4, 16, 1};
+            size_t global_size_s[3] = {static_cast<size_t>(width_s), static_cast<size_t>(height_s), 1};
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_s, local_size_s, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            // copy transposed buffer contents to original buffers
+            CL_CHECK(clEnqueueCopyBuffer(queue, qT_d, extra->q, 0, 0, q_size_bytes, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, dT_d, extra->d, 0, 0, d_size_bytes, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clReleaseMemObject(qT_d));
+            CL_CHECK(clReleaseMemObject(dT_d));
+
+            CL_CHECK(clReleaseMemObject(q_d_image1D));
+            CL_CHECK(clReleaseMemObject(d_d_image1D));
+            CL_CHECK(clReleaseMemObject(qT_d_image1D));
+            CL_CHECK(clReleaseMemObject(dT_d_image1D));
+        } // end transpose
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl * extra_orig = (ggml_tensor_extra_cl *)tensor->extra;
+        GGML_ASSERT(extra_orig && "Tesnors in OpenCL backend should have been allocated and initialized");
+
+        // Allocate the new extra and create aliases from the original.
+        ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+        ggml_tensor_extra_cl_q6_K * extra = ctx->ggml_opencl_alloc_temp_tensor_extra_q6_K();
+
+        size_t size_ql = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/2;
+        size_t size_qh = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/4;
+        size_t size_s  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*ggml_blck_size(tensor->type)/16;
+        size_t size_d  = ggml_nelements(tensor)/ggml_blck_size(tensor->type)*sizeof(ggml_fp16_t);
+        GGML_ASSERT(size_ql + size_qh + size_s + size_d == ggml_nbytes(tensor) &&
+            "Incorrect tensor size");
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+        CL_CHECK(clEnqueueWriteBuffer(
+            queue, data_device, CL_TRUE, 0,
+            ggml_nbytes(tensor), data, 0, NULL, NULL));
+
+        cl_buffer_region region;
+
+        // Subbuffer for ql
+        region.origin = align_to(extra_orig->offset + tensor->view_offs + offset, backend_ctx->alignment);
+        region.size = size_ql;
+        extra->ql = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        auto previous_origin = region.origin;
+
+        // Subbuffer for qh
+        region.origin = align_to(previous_origin + size_ql, backend_ctx->alignment);
+        region.size = size_qh;
+        extra->qh = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Subbuffer for scales
+        region.origin = align_to(previous_origin + size_qh, backend_ctx->alignment);
+        region.size = size_s;
+        extra->s = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Create subbuffer for d.
+        region.origin = align_to(previous_origin + size_s, backend_ctx->alignment);
+        region.size = size_d;
+        extra->d = clCreateSubBuffer(
+            extra_orig->data_device, CL_MEM_READ_WRITE,
+            CL_BUFFER_CREATE_TYPE_REGION, &region, &err);
+        CL_CHECK(err);
+        previous_origin = region.origin;
+
+        // Flatten the weights
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q6_K;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra->d));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clReleaseMemObject(data_device));
+
+        extra->size_ql = size_ql;
+        extra->size_qh = size_qh;
+        extra->size_s  = size_s;
+        extra->size_d  = size_d;
+
+        tensor->extra  = extra;
+        return;
+    }
+#endif // GGML_OPENCL_SOA_Q
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+    GGML_ASSERT(extra);
+
+    CL_CHECK(clEnqueueWriteBuffer(
+        queue, extra->data_device, CL_TRUE, extra->offset + offset,
+        size, data, 0, NULL, NULL));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->extra);
+
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(buffer->buft->device);
+
+    cl_context context = backend_ctx->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    // Make sure all previously submitted commands in other devices are finished.
+    sync_with_other_backends(backend_ctx);
+
+#ifdef GGML_OPENCL_SOA_Q
+    // In end-to-end runs, get_tensor is usually used to get back the logits,
+    // where we can simply do clEnqueueReadBuffer since they are f32.
+    // However, in test-backend-ops, the GPU graph is copied to the CPU backend,
+    // which requires reading back quantized weight tensors.
+    // To properly support this, we need to restore block_q4_0 struct arrays
+    // from the flattened buffers.
+    if (tensor->type == GGML_TYPE_Q4_0) {
+        ggml_tensor_extra_cl_q4_0 * extra = (ggml_tensor_extra_cl_q4_0 *)tensor->extra;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            cl_int err;
+            cl_kernel kernel;
+
+            cl_int M = tensor->ne[1];   // ne01
+            cl_int K = tensor->ne[0];   // ne00
+
+            GGML_ASSERT(K % 32 == 0);
+            GGML_ASSERT(M % 4 == 0);
+
+            size_t size_q = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*ggml_blck_size(tensor->type)/2;
+            size_t size_d = (ggml_nelements(tensor)/ggml_blck_size(tensor->type))*sizeof(ggml_fp16_t);
+            GGML_ASSERT(size_d + size_q == ggml_nbytes(tensor) && "Incorrect tensor size");
+
+            cl_mem buf_trans_q;
+            cl_mem buf_trans_d;
+
+            CL_CHECK((buf_trans_q = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                size_q, NULL, &err), err));
+            CL_CHECK((buf_trans_d = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                size_d, NULL, &err), err));
+
+            kernel = backend_ctx->kernel_transpose_16_buf;
+
+            // transpose q back
+            cl_int stride_k_q = K/4;
+            size_t local_size_q[3] = {64, 1, 1};
+            size_t global_size_q[3] = {(size_t)M, (size_t)stride_k_q, 1};
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_q));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_int), &M));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &stride_k_q));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_size_q, local_size_q, 0, NULL, NULL));
+
+            // transpose scales back
+            cl_int stride_k_d = K/32;
+            size_t local_size_d[3] = {64, 1, 1};
+            size_t global_size_d[3] = {(size_t)M, (size_t)stride_k_d, 1};
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->d));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &buf_trans_d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_int), &M));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &stride_k_d));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_size_d, local_size_d, 0, NULL, NULL));
+
+            // unpack
+            cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+                ggml_nbytes(tensor), NULL, &err);
+            CL_CHECK(err);
+
+            cl_uchar mask_0F = 0x0F;
+            cl_uchar mask_F0 = 0xF0;
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            kernel = backend_ctx->kernel_restore_block_q4_0_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_uchar), &mask_F0));
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, NULL));
+
+            // read back to host
+            CL_CHECK(clEnqueueReadBuffer(
+                queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+
+            CL_CHECK(clReleaseMemObject(data_device));
+            CL_CHECK(clReleaseMemObject(buf_trans_q));
+            CL_CHECK(clReleaseMemObject(buf_trans_d));
+
+            return;
+        }
+#endif
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q4_0;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_moe_kernels(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_restore_block_mxfp4_trans;
+
+            int ne00 = tensor->ne[0];
+            int ne01 = tensor->ne[1];
+            int ne02 = tensor->ne[2];
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->e));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &ne00));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_int), &ne01));
+
+            size_t global_work_size[3] = {static_cast<size_t>(((ne01 + 63) / 64) * 64), static_cast<size_t>(ne00 / 32), static_cast<size_t>(ne02)};
+            size_t local_work_size[3] = {64, 2, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+            CL_CHECK(clEnqueueReadBuffer(
+                queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+            CL_CHECK(clReleaseMemObject(data_device));
+            return;
+        }
+#endif
+        cl_kernel kernel = backend_ctx->kernel_restore_block_mxfp4;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->e));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q8_0) {
+        ggml_tensor_extra_cl_q8_0 * extra = (ggml_tensor_extra_cl_q8_0 *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (enable_adreno_trans_weight(backend_ctx, tensor)) {
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0_trans;
+
+            int ne00 = tensor->ne[0];
+            int ne01 = tensor->ne[1];
+            GGML_ASSERT(tensor->ne[2] == 1);  // ???
+            GGML_ASSERT(tensor->ne[3] == 1);  // ???
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_int), &ne00));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_int), &ne01));
+
+            size_t global_work_size[3] = {static_cast<size_t>(((ne01 + 63) / 64) * 64), 1, 1};
+            size_t local_work_size[3] = {64, 1, 1};
+
+            cl_event evt;
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, &evt));
+            CL_CHECK(clWaitForEvents(1, &evt));
+
+            CL_CHECK(clEnqueueReadBuffer(
+                queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+            CL_CHECK(clReleaseMemObject(data_device));
+            return;
+        }
+#endif
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q8_0;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->q));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+    if (tensor->type == GGML_TYPE_Q6_K) {
+        ggml_tensor_extra_cl_q6_K * extra = (ggml_tensor_extra_cl_q6_K *)tensor->extra;
+
+        cl_int err;
+        cl_mem data_device = clCreateBuffer(context, CL_MEM_READ_WRITE,
+            ggml_nbytes(tensor), NULL, &err);
+        CL_CHECK(err);
+
+        cl_kernel kernel = backend_ctx->kernel_restore_block_q6_K;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra->ql));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra->qh));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra->s));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &extra->d));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &data_device));
+
+        size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+        size_t local_work_size[] = {1, 1, 1};
+
+        cl_event evt;
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+            global_work_size, local_work_size, 0, NULL, &evt));
+        CL_CHECK(clWaitForEvents(1, &evt));
+        CL_CHECK(clEnqueueReadBuffer(
+            queue, data_device, CL_TRUE, offset,
+            size, data, 0, NULL, NULL));
+        CL_CHECK(clReleaseMemObject(data_device));
+        return;
+    }
+#endif // GGML_OPENCL_SOA_Q
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+
+    CL_CHECK(clEnqueueReadBuffer(
+        queue, extra->data_device, CL_TRUE, extra->offset + tensor->view_offs + offset,
+        size, data, 0, NULL, NULL));
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_opencl_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_dev_t dev = buffer->buft->device;
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(dev);
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    for (cl_mem buf : ctx->buffer) {
+        CL_CHECK(clEnqueueFillBuffer(queue, buf, &value, sizeof(value), 0, buffer->size, 0, NULL, NULL));
+    }
+    CL_CHECK(clFinish(queue));
+}
+
+static void ggml_backend_opencl_buffer_reset(ggml_backend_buffer_t buffer) {
+    ggml_backend_opencl_buffer_context * ctx = (ggml_backend_opencl_buffer_context *) buffer->context;
+    ctx->reset();
+}
+
+static ggml_backend_buffer_i ggml_backend_opencl_buffer_interface = {
+    /* .free_buffer     = */ ggml_backend_opencl_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_opencl_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_opencl_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_opencl_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_opencl_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_opencl_buffer_clear,
+    /* .reset           = */ ggml_backend_opencl_buffer_reset,
+};
+
+//
+// buffer type
+//
+
+static const char * ggml_backend_opencl_buffer_type_get_name(ggml_backend_buffer_type_t buffer_type) {
+    return "OpenCL";
+
+    GGML_UNUSED(buffer_type);
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buffer_type, size_t size) {
+    ggml_backend_opencl_context *backend_ctx = ggml_cl2_init(buffer_type->device);
+
+    // clCreateBuffer returns -61 for size 0
+    size = std::max(size, (size_t)1);
+
+    cl_int err;
+    cl_mem mem = clCreateBuffer(backend_ctx->context, CL_MEM_READ_WRITE, size, NULL, &err);
+    if (err != CL_SUCCESS) {
+        GGML_LOG_INFO("%s: failed to allocate %.2f MiB\n", __func__, size / 1024.0 / 1024.0);
+        return nullptr;
+    }
+
+    ggml_backend_opencl_buffer_context * ctx = new ggml_backend_opencl_buffer_context(mem);
+
+    return ggml_backend_buffer_init(buffer_type, ggml_backend_opencl_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_alignment(ggml_backend_buffer_type_t buffer_type) {
+    ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer_type->device);
+    return backend_ctx->alignment;
+}
+
+static size_t ggml_backend_opencl_buffer_type_get_max_size(ggml_backend_buffer_type_t buffer_type) {
+    static size_t max_size = -1;
+    if (max_size == (size_t)-1) {
+        ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(buffer_type->device);
+        max_size = backend_ctx->max_alloc_size;
+    }
+    return max_size;
+}
+
+static bool ggml_backend_opencl_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
+    return ggml_backend_is_opencl(backend);
+
+    UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_opencl_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_opencl_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_opencl_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_opencl_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_opencl_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ NULL,
+    /* .is_host          = */ NULL,
+};
+
+//
+// backend device
+//
+
+static const char * ggml_backend_opencl_device_get_name(ggml_backend_dev_t dev) {
+    return "GPUOpenCL";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_opencl_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_opencl_device_context *dev_ctx = (ggml_backend_opencl_device_context *) dev->context;
+    return dev_ctx->device_name.c_str();
+}
+
+static void ggml_backend_opencl_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_opencl_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_GPU;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_opencl_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_opencl_device_get_name(dev);
+    props->description = ggml_backend_opencl_device_get_description(dev);
+    props->type        = ggml_backend_opencl_device_get_type(dev);
+    ggml_backend_opencl_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = ggml_backend_dev_caps {
+        /* .async                 = */ false,
+        /* .host_buffer           = */ false,
+        /* .buffer_from_host_ptr  = */ false,
+        /* .events                = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_opencl_device_init(ggml_backend_dev_t dev, const char * params) {
+    ggml_backend_opencl_context * backend_ctx = ggml_cl2_init(dev);
+    // Getting a new reference to the backend, increase ref_count
+    backend_ctx->ref_count++;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_opencl_guid(),
+        /* .interface = */ ggml_backend_opencl_i,
+        /* .device    = */ dev,
+        /* .context   = */ backend_ctx,
+    };
+
+    return backend;
+
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_opencl_device_get_buffer_type(ggml_backend_dev_t dev) {
+    auto * dev_ctx = static_cast<ggml_backend_opencl_device_context *>(dev->context);
+
+    dev_ctx->buffer_type = ggml_backend_buffer_type{
+        /* .iface   = */ ggml_backend_opencl_buffer_type_interface,
+        /* .device  = */ dev,
+        /* .context = */ nullptr,
+    };
+
+    return &dev_ctx->buffer_type;
+}
+
+static ggml_backend_buffer_t ggml_backend_opencl_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    GGML_UNUSED(dev);
+    GGML_UNUSED(ptr);
+    GGML_UNUSED(size);
+    GGML_UNUSED(max_tensor_size);
+    return nullptr;
+}
+
+static bool ggml_backend_opencl_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    return ggml_opencl_supports_op(dev, op);
+}
+
+static bool ggml_backend_opencl_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    // Check 'dev' and 'buffer_type' are not objects belonging to this backend.
+    if (dev->iface.get_name != ggml_backend_opencl_device_get_name ||
+        buft->iface.get_name != ggml_backend_opencl_buffer_type_get_name) {
+        return false;
+    }
+
+    // Check cl_context is the same. clEnqueue* commands may not use
+    // buffers from another cl_context.
+    ggml_backend_opencl_context * backend_ctx0 = ggml_cl2_init(dev);
+    ggml_backend_opencl_context * backend_ctx1 = ggml_cl2_init(buft->device);
+    return backend_ctx0->context == backend_ctx1->context;
+}
+
+namespace /* anonymous */ {
+struct ggml_backend_device_i ggml_backend_opencl_device_i = {
+    /* .get_name             = */ ggml_backend_opencl_device_get_name,
+    /* .get_description      = */ ggml_backend_opencl_device_get_description,
+    /* .get_memory           = */ ggml_backend_opencl_device_get_memory,
+    /* .get_type             = */ ggml_backend_opencl_device_get_type,
+    /* .get_props            = */ ggml_backend_opencl_device_get_props,
+    /* .init_backend         = */ ggml_backend_opencl_device_init,
+    /* .get_buffer_type      = */ ggml_backend_opencl_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_opencl_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_opencl_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_opencl_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+}
+
+// Backend registry
+
+static const char * ggml_backend_opencl_reg_get_name(ggml_backend_reg_t reg) {
+    return "OpenCL";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_opencl_reg_device_count(ggml_backend_reg_t reg) {
+    return g_ggml_backend_opencl_devices.size();
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_opencl_reg_device_get(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index < ggml_backend_opencl_reg_device_count(reg));
+
+    return &g_ggml_backend_opencl_devices[index];
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static struct ggml_backend_reg_i ggml_backend_opencl_reg_i = {
+    /* .get_name         = */ ggml_backend_opencl_reg_get_name,
+    /* .device_count     = */ ggml_backend_opencl_reg_device_count,
+    /* .device_get       = */ ggml_backend_opencl_reg_device_get,
+    /* .get_proc_address = */ NULL,
+};
+
+ggml_backend_reg_t ggml_backend_opencl_reg(void) {
+    static std::mutex mutex;
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    if (initialized) {
+        return &reg;
+    }
+    initialized = true;
+
+    g_ggml_backend_opencl_devices = ggml_opencl_probe_devices(&reg);
+
+    reg = ggml_backend_reg{
+        /* .api_version = */ GGML_BACKEND_API_VERSION,
+        /* .iface       = */ ggml_backend_opencl_reg_i,
+        /* .context     = */ NULL,
+    };
+
+    return &reg;
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_opencl_reg)
+
+//------------------------------------------------------------------------------
+// Debugging utils
+//------------------------------------------------------------------------------
+#if 0
+#define QK4_0 32
+typedef struct {
+    ggml_fp16_t d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2,
+    "wrong q4_0 block size/padding");
+
+#include <math.h>
+#ifdef __cplusplus
+#include "half.hpp"
+#endif
+
+static void dump_tensor(ggml_backend_t backend, const struct ggml_tensor * tensor) {
+    void * buf = malloc(ggml_nbytes(tensor));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+#ifdef GGML_OPENCL_SOA_Q
+    void * buf_q;
+    void * buf_d;
+#endif
+
+    // Make sure everything is done.
+    CL_CHECK(clFinish(queue));
+
+#ifdef GGML_OPENCL_SOA_Q
+    if (tensor->type == GGML_TYPE_Q4_0) {
+        ggml_tensor_extra_cl_q4_0 * extra = (ggml_tensor_extra_cl_q4_0 *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        size_t size_q = ggml_nelements(tensor)/QK4_0 * QK4_0/2;
+        size_t size_d = ggml_nelements(tensor)/QK4_0 * sizeof(ggml_fp16_t);
+        GGML_ASSERT(size_q + size_d == ggml_nbytes(tensor));
+        buf_q = malloc(size_q);
+        buf_d = malloc(size_d);
+
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_d, buf_d, 0, NULL, NULL));
+        CL_CHECK(clFinish(queue));
+    } else if (tensor->type == GGML_TYPE_MXFP4) {
+        ggml_tensor_extra_cl_mxfp4 * extra = (ggml_tensor_extra_cl_mxfp4 *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        size_t size_q = ggml_nelements(tensor)/QK_MXFP4 * QK_MXFP4/2;
+        size_t size_e = ggml_nelements(tensor)/QK_MXFP4 * sizeof(char);
+        GGML_ASSERT(size_q + size_e == ggml_nbytes(tensor));
+        buf_q = malloc(size_q);
+        buf_d = malloc(size_e);
+
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->q, CL_TRUE, 0, size_q, buf_q, 0, NULL, NULL));
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->d, CL_TRUE, 0, size_e, buf_d, 0, NULL, NULL));
+        CL_CHECK(clFinish(queue));
+    } else {
+        // Read out the tensor from GPU memory.
+        ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+        GGML_ASSERT(extra);
+
+        CL_CHECK(clEnqueueReadBuffer(queue, extra->data_device, CL_TRUE,
+        extra->offset, ggml_nbytes(tensor), buf, 0, NULL, NULL));
+        CL_CHECK(clFinish(queue));
+    }
+#else
+    // Read out the tensor from GPU memory.
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *) tensor->extra;
+    GGML_ASSERT(extra);
+
+    CL_CHECK(clEnqueueReadBuffer(queue, extra->data_device, CL_TRUE,
+        extra->offset, ggml_nbytes(tensor), buf, 0, NULL, NULL));
+    CL_CHECK(clFinish(queue));
+#endif // GGML_OPENCL_SOA_Q
+
+    // Open file and dump.
+    char fname[512];
+    snprintf(fname, sizeof(fname), "./tensor-dumps/%s.txt", tensor->name);
+    FILE * f = fopen(fname, "w");
+    if (!f) {
+        printf("Failed to open %s\n", fname);
+        return;
+    }
+
+    if (tensor->type == GGML_TYPE_F32) {
+        float * data = (float *) buf;
+        for (int i = 0; i < ggml_nelements(tensor); ++i) {
+            if (isnan(data[i])) {
+                printf("NaN found: %s\n", tensor->name);
+                break;
+            }
+            fprintf(f, "%f\n", data[i]);
+        }
+    } else if (tensor->type == GGML_TYPE_I32) {
+        int * data = (int *) buf;
+        for (int i = 0; i < ggml_nelements(tensor); ++i) {
+            if (isnan(data[i])) {
+                printf("NaN found: %s\n", tensor->name);
+                break;
+            }
+            fprintf(f, "%d\n", data[i]);
+        }
+    } else if (tensor->type == GGML_TYPE_F16) {
+#ifdef __cplusplus
+        half_float::half * data = (half_float::half *) buf;
+        for (int i = 0; i < ggml_nelements(tensor); ++i) {
+            if (std::isnan(data[i])) {
+                printf("NaN found: %s\n", tensor->name);
+                break;
+            }
+            fprintf(f, "%f\n", float(data[i]));
+        }
+#endif
+    } else if (tensor->type == GGML_TYPE_Q4_0) {
+#ifdef GGML_OPENCL_SOA_Q
+        ggml_fp16_t * data_d = (ggml_fp16_t *)buf_d;
+        unsigned char * data_q = (unsigned char *)buf_q;
+
+        for (int i = 0; i < ggml_nelements(tensor)/QK4_0; ++i) {
+            fprintf(f, "%04x, ", data_d[i]);
+            for (int k = 0; k < QK4_0/2; ++k) {
+                fprintf(f, "%02x, ", data_q[k]);
+            }
+            fprintf(f, "\n");
+            data_q += QK4_0/2;
+        }
+        free(buf_d);
+        free(buf_q);
+#else
+        block_q4_0 * data = (block_q4_0 *) buf;
+        for (int i = 0; i < ggml_nelements(tensor)/QK4_0; ++i) {
+            fprintf(f, "%04x, ", data[i].d);
+            for (int k = 0; k < QK4_0/2; ++k) {
+                fprintf(f, "%02x, ", data[i].qs[k]);
+            }
+            fprintf(f, "\n");
+        }
+#endif // GGML_OPENCL_SOA_Q
+    }
+    free(buf);
+    fflush(f);
+    fclose(f);
+}
+#else
+#define dump_tensor(tensor)
+#endif
+
+//------------------------------------------------------------------------------
+// Ops
+//------------------------------------------------------------------------------
+
+static bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
+            src1->type == GGML_TYPE_F32 &&
+             dst->type == GGML_TYPE_F32 &&
+            (ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
+}
+
+// Copy a noncontiguous tensor to contiguous tensor. ne[] remains the same but
+// nb[] is recalculated such that tensor is contiguous.
+static void ggml_cl_copy_to_contiguous(ggml_backend_t backend, const ggml_tensor * src, cl_mem dst,
+                                       cl_ulong &nb0, cl_ulong &nb1, cl_ulong &nb2, cl_ulong &nb3) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int tensor_type_size = ggml_type_size(src->type);
+
+    const int ne00 = src->ne[0];
+    const int ne01 = src->ne[1];
+    const int ne02 = src->ne[2];
+    const int ne03 = src->ne[3];
+
+    const cl_ulong nb00 = src->nb[0];
+    const cl_ulong nb01 = src->nb[1];
+    const cl_ulong nb02 = src->nb[2];
+    const cl_ulong nb03 = src->nb[3];
+
+    const int ne0 = src->ne[0];
+    const int ne1 = src->ne[1];
+    const int ne2 = src->ne[2];
+    const int ne3 = src->ne[3];
+
+    nb0 = tensor_type_size;
+    nb1 = tensor_type_size*ne00;
+    nb2 = tensor_type_size*ne00*ne01;
+    nb3 = tensor_type_size*ne00*ne01*ne02;
+
+    ggml_tensor_extra_cl * extra = (ggml_tensor_extra_cl *)src->extra;
+
+    cl_ulong offset0 = extra->offset + src->view_offs;
+    cl_ulong offsetd = 0;
+
+    cl_kernel kernel;
+
+    switch (src->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_cpy_f32_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_cpy_f16_f16;
+            break;
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &dst));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+
+    const int nth = MIN(64, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src);
+}
+
+static void ggml_cl_nop(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    UNUSED(backend);
+    UNUSED(src0);
+    UNUSED(src1);
+    UNUSED(dst);
+}
+
+static void ggml_cl_get_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const int      ne00 = src0->ne[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+    const int      ne10 = src1->ne[0];
+    const cl_ulong nb10 = src1->nb[0];
+    const int      ne11 = src1->ne[1];
+    const int      ne12 = src1->ne[2];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            kernel = backend_ctx->kernel_get_rows_f32;
+            break;
+        case GGML_TYPE_F16:
+            kernel = backend_ctx->kernel_get_rows_f16;
+            break;
+        case GGML_TYPE_Q4_0:
+            kernel = backend_ctx->kernel_get_rows_q4_0;
+            break;
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb3));
+
+    size_t global_work_size[] = {(size_t)ne10*64, (size_t)ne11, (size_t)ne12};
+    size_t local_work_size[] = {64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_set_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src1->type == GGML_TYPE_I64 || src1->type == GGML_TYPE_I32);
+
+    // ne0 = ne00
+    // ne2 = ne02
+    // ne3 = ne03
+
+    const int      ne01 = src0->ne[1];
+    const int      ne02 = src0->ne[2];
+    const int      ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int      ne11 = src1->ne[1];
+    const int      ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+
+    const int      ne0  = dst->ne[0];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    const int nblk0 = ne0/ggml_blck_size(dst->type);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f32_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f32_i32;
+            }
+            break;
+        case GGML_TYPE_F16:
+            if (src1->type == GGML_TYPE_I64) {
+                kernel = backend_ctx->kernel_set_rows_f16_i64;
+            } else {
+                kernel = backend_ctx->kernel_set_rows_f16_i32;
+            }
+            break;
+        default:
+            GGML_ABORT("not implemented");
+    }
+
+    fastdiv_vals ne11_ = init_fastdiv_values(ne11);
+    fastdiv_vals ne12_ = init_fastdiv_values(ne12);
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(fastdiv_vals), &ne11_));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(fastdiv_vals), &ne12_));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &nblk0));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb3));
+
+    int nth0 = 64;
+    if (backend_ctx->gpu_family == INTEL) {
+        nth0 = 32;
+    } else if (backend_ctx->gpu_family == ADRENO) {
+        nth0 = 64;
+    }
+
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth0 < nblk0 && nth0 < max_workgroup_size) {
+        nth0 *= 2;
+    }
+
+    int rows_per_workgroup = 1;
+    if (nth0 > nblk0) {
+        rows_per_workgroup = nth0 / nblk0;
+        nth0 = nblk0;
+    }
+
+    size_t global_work_size[] = {
+        (size_t)(ne01 + rows_per_workgroup - 1)/rows_per_workgroup*nth0,
+        (size_t)ne02*rows_per_workgroup,
+        (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth0, (size_t)rows_per_workgroup, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_add(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    const bool bcast_row = ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0;
+
+    if (bcast_row) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+        GGML_ASSERT(ne11 == 1);
+    }
+
+    if (dst->type == GGML_TYPE_F32) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+        } else {
+            kernel = backend_ctx->kernel_add;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        GGML_ASSERT(src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F32);
+        GGML_ASSERT(src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+        const int type_src0 = (src0->type == GGML_TYPE_F32);
+        const int type_src1 = (src1->type == GGML_TYPE_F32);
+        if (bcast_row) {
+            kernel = backend_ctx->kernel_add_row_f16;
+            const int ne = ne00 / 4;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),      &type_src1));
+        } else {
+            kernel = backend_ctx->kernel_add_f16;
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+            CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+            CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+            CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+            CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+            CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+            CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &type_src0));
+            CL_CHECK(clSetKernelArg(kernel, 31, sizeof(int),      &type_src1));
+        }
+    } else {
+        GGML_ASSERT(false && "unsupported data types for add");
+    }
+
+    if (bcast_row) {
+        int n = ggml_nelements(dst)/4;
+        size_t global_work_size[] = {(size_t)n, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;
+        }
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size_ptr, dst);
+    } else {
+        unsigned int nth = MIN(64, ne0);
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_add_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src2->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(src0));
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const cl_ulong nb11 = src1->nb[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel = backend_ctx->kernel_add_id;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne1));
+
+    int nth = MIN(ne00, (int) backend_ctx->get_kernel_workgroup_size(kernel));
+    size_t global_work_size[] = { (size_t)ne01*nth, (size_t)ne02, 1 };
+    size_t local_work_size[] = { (size_t)nth, 1, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_mul(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3]; UNUSED(ne13);
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3]; UNUSED(nb13);
+
+    const int ne0  = dst->ne[0];
+    const int ne1  = dst->ne[1];
+    const int ne2  = dst->ne[2];
+    const int ne3  = dst->ne[3];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    bool bcast_row = false;
+    cl_kernel kernel;
+
+    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+
+        // src1 is a row
+        GGML_ASSERT(ne11 == 1);
+
+        bcast_row = true;
+        int ne = ne00 / 4;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul_row;
+        } else {
+            kernel = backend_ctx->kernel_mul_row_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_mul;
+        } else {
+            kernel = backend_ctx->kernel_mul_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne13));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &ne3));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nb3));
+    }
+
+    if (bcast_row) {
+        int n = ggml_nelements(dst)/4;
+        size_t global_work_size[] = {(size_t)n, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+        }
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+    } else {
+        unsigned int nth = MIN(64, ne0);
+        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_div(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne0  = dst->ne[0];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    bool bcast_row = false;
+    cl_kernel kernel;
+
+    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+
+        // src1 is a row
+        GGML_ASSERT(ne11 == 1);
+
+        bcast_row = true;
+        int ne = ne00 / 4;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div_row;
+        } else {
+            kernel = backend_ctx->kernel_div_row_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_div;
+        } else {
+            kernel = backend_ctx->kernel_div_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb13));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
+    }
+
+    if (bcast_row) {
+        int n = ggml_nelements(dst)/4;
+        size_t global_work_size[] = {(size_t)n, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        unsigned int nth = MIN(64, ne0);
+        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_sub(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne0  = dst->ne[0];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    bool bcast_row = false;
+    cl_kernel kernel;
+
+    if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+        GGML_ASSERT(ggml_is_contiguous(src0));
+
+        // src1 is a row
+        GGML_ASSERT(ne11 == 1);
+
+        bcast_row = true;
+        int ne = ne00 / 4;
+
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub_row;
+        } else {
+            kernel = backend_ctx->kernel_sub_row_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne));
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sub;
+        } else {
+            kernel = backend_ctx->kernel_sub_f16;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb13));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
+    }
+
+    if (bcast_row) {
+        int n = ggml_nelements(dst)/4;
+        size_t global_work_size[] = {(size_t)n, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        unsigned int nth = MIN(64, ne0);
+        size_t global_work_size[] = {ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_sqr(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqr_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqr_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_sqrt(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    // Currently assumes src0 is contiguous
+    int n = ggml_nelements(dst);
+    if (n % 4 == 0) {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32_4;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16_4;
+        }
+        n /= 4;
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_sqrt_cont_f32;
+        } else {
+            kernel = backend_ctx->kernel_sqrt_cont_f16;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_mean(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_mean_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb3));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_ssm_conv(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    int ne01 = src0->ne[1];
+    cl_ulong nb00 = src0->nb[0];
+    cl_ulong nb01 = src0->nb[1];
+    cl_ulong nb02 = src0->nb[2];
+
+    int ne10 = src1->ne[0];
+    cl_ulong nb11 = src1->nb[1];
+
+    int ne1  = dst->ne[1];
+    int ne2  = dst->ne[2];
+    cl_ulong nb0 = dst->nb[0];
+    cl_ulong nb1 = dst->nb[1];
+    cl_ulong nb2 = dst->nb[2];
+
+    cl_kernel kernel = backend_ctx->kernel_ssm_conv_f32_f32;
+
+    if (ne10 % 4 == 0) {
+        kernel = backend_ctx->kernel_ssm_conv_f32_f32_4;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb2));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne1, (size_t)ne2};
+    size_t local_work_size[]  = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (ne01 % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_gelu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_gelu_erf(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_erf_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_erf;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_gelu_quick(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_gelu_quick_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_gelu_quick;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_silu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_silu_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_silu;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_relu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel = backend_ctx->kernel_relu;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    const int64_t n = ggml_nelements(dst);
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_sigmoid_f32;
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_sigmoid_f16;
+    } else {
+        GGML_ASSERT(false && "Unsupported data types for sigmoid (input and output must be both f32 or f16)");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+    const int64_t n = ggml_nelements(dst);
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_tri(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int tri_type = ggml_get_op_params_i32(dst, 0);
+    const int64_t n = ggml_nelements(dst);
+    const int     ne0  = dst->ne[0];
+    const int     ne1  = dst->ne[1];
+
+    cl_kernel kernel = backend_ctx->kernel_tri;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &n));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &tri_type));
+
+    size_t local_work_size[1] = { 256 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_fill(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src0);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float v = 0.0f;
+    memcpy(&v, ((int32_t *) dst->op_params), sizeof(float));
+
+    const int64_t n = ggml_nelements(dst);
+
+    cl_kernel kernel = backend_ctx->kernel_fill;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(float),    &v));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(float),    &n));
+
+    size_t local_work_size[1] = { 256 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float min;
+    float max;
+    memcpy(&min, ((int32_t *) dst->op_params) + 0, sizeof(float));
+    memcpy(&max, ((int32_t *) dst->op_params) + 1, sizeof(float));
+
+    cl_kernel kernel = backend_ctx->kernel_clamp;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &min));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(float),    &max));
+
+    const int64_t n = ggml_nelements(dst);
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    const int ne00 = src0 ? src0->ne[0] : 0;
+    const int ne01 = src0 ? src0->ne[1] : 0;
+    const int ne02 = src0 ? src0->ne[2] : 0;
+    const int ne03 = src0 ? src0->ne[3] : 0;
+
+    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
+    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
+    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+
+    const int nth = MIN(64, ne00);
+
+    cl_kernel kernel = backend_ctx->kernel_norm;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),    &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong),  &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),    &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong),  &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),       &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),       &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),       &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),       &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),     &eps));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth, NULL));
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_rms_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    //ggml_backend_opencl_device_context * dev_ctx =
+    //    (ggml_backend_opencl_device_context *)backend->device->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    const int ne00 = src0 ? src0->ne[0] : 0;
+    const int ne01 = src0 ? src0->ne[1] : 0;
+    const int ne02 = src0 ? src0->ne[2] : 0;
+    const int ne03 = src0 ? src0->ne[3] : 0;
+
+    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
+    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
+    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+
+    GGML_ASSERT(ne00 % 4 == 0);
+
+    const int nth = MIN(64, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    cl_kernel kernel = backend_ctx->kernel_rms_norm;
+
+    // Note, this kernel declares local memory in kernel args and the size
+    // depends on subgroup size.
+    // Note, this requires OpenCL 2.1 and above
+    // For now we use fixed subgroup size to simplify support for OpenCL 2.0.
+    size_t sgs;
+    //CL_CHECK(clGetKernelSubGroupInfo(kernel, dev_ctx->device,
+    //    CL_KERNEL_MAX_SUB_GROUP_SIZE_FOR_NDRANGE,
+    //    sizeof(local_work_size), local_work_size,
+    //    sizeof(size_t), &sgs, NULL));
+    if (backend_ctx->gpu_family == ADRENO) {
+        sgs = 64;
+    } else if (backend_ctx->gpu_family == INTEL) {
+        sgs = 32;
+    } else {
+        GGML_ASSERT(false && "Unsupported GPU");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),    &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong),  &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),    &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong),  &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),       &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),       &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),       &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),       &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(float),     &eps));
+    // This is local memory - the size depends on subgroup size.
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float)*nth/sgs,  NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_opencl_op_rms_norm_fused(ggml_backend_t backend, ggml_tensor * rms_norm_tensor, ggml_tensor * mul_tensor) {
+    GGML_ASSERT(mul_tensor);
+    GGML_ASSERT(rms_norm_tensor);
+
+    // src0 is the src of rms_norm, src1 is the other src of mul (one being rms_norm)
+    const ggml_tensor * src0 = rms_norm_tensor->src[0];
+    const ggml_tensor * src1;
+    if (mul_tensor->src[0] == rms_norm_tensor) {
+        src1 = mul_tensor->src[1];
+    } else if (mul_tensor->src[1] == rms_norm_tensor) {
+        src1 = mul_tensor->src[0];
+    } else {
+        GGML_ASSERT(false && "Invalid args for rms_norm and mul");
+    }
+    const ggml_tensor * dst = mul_tensor;
+
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float eps;
+    memcpy(&eps, rms_norm_tensor->op_params, sizeof(float));
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    GGML_ASSERT(ne00 % 4 == 0);
+
+    size_t sgs;
+    if (backend_ctx->gpu_family == ADRENO) {
+        sgs = 64;
+    } else if (backend_ctx->gpu_family == INTEL) {
+        sgs = 32;
+    } else {
+        GGML_ASSERT(false && "Unsupported GPU");
+    }
+
+    cl_kernel kernel = backend_ctx->kernel_rms_norm_mul;
+
+    int nth = sgs;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth < ne00 && nth < max_workgroup_size) {
+        nth *= 2;
+    }
+    nth = MIN(nth, max_workgroup_size);
+    nth = MIN(nth, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),        &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong),      &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),        &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong),      &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),        &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong),      &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),           &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),           &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),           &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),           &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),      &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),      &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),      &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),           &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),           &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),           &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),           &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),      &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),      &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),      &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong),      &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong),      &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),      &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),         &eps));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,     NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_opencl_op_norm_fused(ggml_backend_t backend, ggml_tensor * norm_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(norm_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = norm_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == norm_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float eps;
+    memcpy(&eps, norm_tensor->op_params, sizeof(float));
+
+    const int ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+    const cl_ulong nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+    const int ne10 = src1->ne[0], ne11 = src1->ne[1], ne12 = src1->ne[2], ne13 = src1->ne[3];
+    const cl_ulong nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+    const int ne20 = src2->ne[0], ne21 = src2->ne[1], ne22 = src2->ne[2], ne23 = src2->ne[3];
+    const cl_ulong nb21 = src2->nb[1], nb22 = src2->nb[2], nb23 = src2->nb[3];
+    const cl_ulong nbd1 = dst->nb[1], nbd2 = dst->nb[2], nbd3 = dst->nb[3];
+
+    size_t sgs;
+    if (backend_ctx->gpu_family == ADRENO) sgs = 64;
+    else if (backend_ctx->gpu_family == INTEL) sgs = 32;
+    else GGML_ASSERT(false && "Unsupported GPU");
+
+    cl_kernel kernel = backend_ctx->kernel_norm_mul_add;
+
+    int nth = sgs;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    while (nth < ne00/4 && nth < max_workgroup_size) nth *= 2;
+    nth = MIN(nth, max_workgroup_size);
+    nth = MIN(nth, ne00/4);
+
+    size_t gws[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t lws[] = {(size_t)nth, 1, 1};
+    size_t num_subgroups = (nth + sgs - 1) / sgs;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int), &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int), &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int), &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int), &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int), &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int), &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int), &ne20));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int), &ne21));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int), &ne22));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int), &ne23));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(cl_ulong), &nb21));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(cl_ulong), &nb22));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(cl_ulong), &nb23));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(cl_ulong), &nbd1));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(cl_ulong), &nbd2));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_ulong), &nbd3));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float), &eps));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_float2) * num_subgroups, NULL));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, gws, lws, dst);
+}
+
+static void ggml_opencl_op_group_norm_fused(ggml_backend_t backend, ggml_tensor * gn_tensor, ggml_tensor * mul_tensor, ggml_tensor * add_tensor) {
+    GGML_ASSERT(gn_tensor && mul_tensor && add_tensor);
+
+    const ggml_tensor * src0 = gn_tensor->src[0];
+    const ggml_tensor * src1 = mul_tensor->src[0] == gn_tensor ? mul_tensor->src[1] : mul_tensor->src[0];
+    const ggml_tensor * src2 = add_tensor->src[0] == mul_tensor ? add_tensor->src[1] : add_tensor->src[0];
+    const ggml_tensor * dst = add_tensor;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    int groups;
+    float eps;
+    memcpy(&groups, gn_tensor->op_params, sizeof(int));
+    memcpy(&eps, (char *)gn_tensor->op_params + sizeof(int), sizeof(float));
+
+    cl_kernel kernel = backend_ctx->kernel_group_norm_mul_add;
+    int max_workgroup_size = backend_ctx->get_kernel_workgroup_size(kernel);
+    int ne = ggml_nelements(src0);
+    int group_size = ne / groups;
+
+    size_t lws[] = { (size_t)MIN(max_workgroup_size, group_size) };
+    size_t gws[] = { (size_t)groups * lws[0] };
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem), &extra2->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem), &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int), &ne));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int), &group_size));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(float), &eps));
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, gws, lws, dst);
+}
+
+static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    int32_t n_groups   = ((const int32_t *) dst->op_params)[0];
+    int32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + n_groups - 1) / n_groups);
+    float   eps        = ((const float *) dst->op_params)[1];
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne = ne00*ne01*ne02;
+
+    cl_kernel kernel = backend_ctx->kernel_group_norm;
+
+    size_t sgs = 64;
+    if (backend_ctx->gpu_family == ADRENO) {
+        sgs = 64;
+    } else if (backend_ctx->gpu_family == INTEL) {
+        sgs = 32;
+    } else {
+        GGML_ASSERT(false && "Unsupported GPU");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &group_size));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(float),    &eps));
+
+    size_t global_work_size[] = {(size_t)n_groups*sgs, 1, 1};
+    size_t local_work_size[] = {(size_t)sgs, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb0  = dst->nb[0];
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    cl_kernel kernel;
+
+    if (ggml_is_contiguous(src0)) {
+        // Handle contiguous input
+        int n = ggml_nelements(dst);
+        if (n % 4 == 0) {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_tanh_f32_4;
+            } else {
+                kernel = backend_ctx->kernel_tanh_f16_4;
+            }
+            n /= 4;
+        } else {
+            if (src0->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_tanh_f32;
+            } else {
+                kernel = backend_ctx->kernel_tanh_f16;
+            }
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+
+        size_t global_work_size[] = {(size_t)n, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        size_t * local_work_size_ptr = local_work_size;
+        if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;
+        }
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+    } else {
+        // Handle non-contiguous input
+        if (src0->type == GGML_TYPE_F32) {
+            kernel = backend_ctx->kernel_tanh_f32_nc;
+        } else {
+            kernel = backend_ctx->kernel_tanh_f16_nc;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &nb00));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb02));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb03));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb0));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb1));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb2));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb3));
+
+        int nth = 64;
+
+        size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+        size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_expm1(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_expm1_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_expm1_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_expm1");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0];
+    const int ne11 = dst->ne[1];
+    const int ne12 = dst->ne[2];
+    const int ne13 = dst->ne[3];
+
+    const cl_ulong nb10 = dst->nb[0];
+    const cl_ulong nb11 = dst->nb[1];
+    const cl_ulong nb12 = dst->nb[2];
+    const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_softplus(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_softplus_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_softplus_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_softplus");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0];
+    const int ne11 = dst->ne[1];
+    const int ne12 = dst->ne[2];
+    const int ne13 = dst->ne[3];
+
+    const cl_ulong nb10 = dst->nb[0];
+    const cl_ulong nb11 = dst->nb[1];
+    const cl_ulong nb12 = dst->nb[2];
+    const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == src0->type);
+
+    UNUSED(src1_shape_def);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd  = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
+
+    const cl_ulong nb0 = dst->nb[0];
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_repeat_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb3));
+
+    int nth = 64;
+
+    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    if (backend_ctx->kernel_pad == nullptr) {
+        GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int s_ne0 = src0->ne[0];
+    const int s_ne1 = src0->ne[1];
+    const int s_ne2 = src0->ne[2];
+    const int s_ne3 = src0->ne[3];
+
+    const int s_nb0 = src0->nb[0];
+    const int s_nb1 = src0->nb[1];
+    const int s_nb2 = src0->nb[2];
+    const int s_nb3 = src0->nb[3];
+
+    const int d_ne0 = dst->ne[0];
+    const int d_ne1 = dst->ne[1];
+    const int d_ne2 = dst->ne[2];
+    const int d_ne3 = dst->ne[3];
+
+    const int d_nb0 = dst->nb[0];
+    const int d_nb1 = dst->nb[1];
+    const int d_nb2 = dst->nb[2];
+    const int d_nb3 = dst->nb[3];
+
+    const int lp0 = ((const int*)(dst->op_params))[0];
+    const int rp0 = ((const int*)(dst->op_params))[1];
+    const int lp1 = ((const int*)(dst->op_params))[2];
+    const int rp1 = ((const int*)(dst->op_params))[3];
+    const int lp2 = ((const int*)(dst->op_params))[4];
+    const int rp2 = ((const int*)(dst->op_params))[5];
+    const int lp3 = ((const int*)(dst->op_params))[6];
+    const int rp3 = ((const int*)(dst->op_params))[7];
+
+    cl_kernel kernel = backend_ctx->kernel_pad;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),       &s_ne0));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),       &s_ne1));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),       &s_ne2));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),       &s_ne3));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),  &s_nb0));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),  &s_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),  &s_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),  &s_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),       &d_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),       &d_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),       &d_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),       &d_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),  &d_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),  &d_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),  &d_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),  &d_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),       &lp0));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),       &rp0));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),       &lp1));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),       &rp1));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),       &lp2));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),       &rp2));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),       &lp3));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(int),       &rp3));
+
+    size_t lws0 = 64;
+    size_t gws0 = (( (size_t)d_ne0 + lws0 - 1 ) / lws0) * lws0;
+
+    size_t global_work_size[] = { gws0, (size_t)d_ne1, (size_t)d_ne2*d_ne3 };
+    size_t local_work_size[]  = { lws0, 1, 1 };
+
+    size_t * local_work_size_ptr = local_work_size;
+     if (d_ne0 % lws0 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int mode_flags        = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    const ggml_scale_mode mode  = (ggml_scale_mode) (mode_flags & 0xFF);
+    cl_kernel kernel = nullptr;
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        kernel = backend_ctx->kernel_upscale;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: nearest upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        kernel = backend_ctx->kernel_upscale_bilinear;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: bilinear upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else {
+        GGML_LOG_WARN("%s: unsupported upscale mode %d, skipping OpenCL execution.\n", __func__, mode);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
+
+    float sf0 = (float)ne0 / ne00;
+    float sf1 = (float)ne1 / ne01;
+    float sf2 = (float)ne2 / ne02;
+    float sf3 = (float)ne3 / ne03;
+
+    float pixel_offset = 0.5f;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne3));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        if (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) {
+            sf0 = ne0 > 1 && ne00 > 1 ? (float)(ne0 - 1) / (ne00 - 1) : sf0;
+            sf1 = ne1 > 1 && ne01 > 1 ? (float)(ne1 - 1) / (ne01 - 1) : sf1;
+            pixel_offset = 0.0f;
+        }
+
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne2));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne3));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float),    &pixel_offset));
+    }
+
+
+    size_t dst_total_elements = (size_t)ne0 * ne1 * ne2 * ne3;
+    if (dst_total_elements == 0) {
+        return;
+    }
+    size_t global_work_size[] = { dst_total_elements, 1, 1 };
+    size_t local_work_size_pref = 256;
+    size_t local_work_size[] = { MIN(local_work_size_pref, dst_total_elements), 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (dst_total_elements % local_work_size[0] != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd  = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+    const int ne2 = dst->ne[2];
+    const int ne3 = dst->ne[3];
+
+    const cl_ulong nb0 = dst->nb[0];
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    const cl_int dim = ((const int32_t *) dst->op_params)[0];
+    GGML_ASSERT(dim >= 0 && dim <= 3);
+
+    int nth = MIN(64, ne0);
+
+    cl_kernel kernel = backend_ctx->kernel_concat_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_int),   &dim));
+
+    size_t global_work_size[] = {(size_t)ne1*nth, (size_t)ne2, (size_t)ne3};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    if (backend_ctx->kernel_timestep_embedding == nullptr) {
+        GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int logical_dim = dst->op_params[0];
+    const int max_period  = dst->op_params[1];
+    const int dst_nb1_bytes = dst->nb[1];
+
+    cl_kernel kernel = backend_ctx->kernel_timestep_embedding;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &dst_nb1_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &logical_dim));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &max_period));
+
+    size_t gws0 = (size_t)(((logical_dim + 1) / 2) + 1);
+
+    size_t gws1 = (size_t)src0->ne[0];
+
+    size_t global_work_size[] = {gws0, gws1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
+}
+
+static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    const ggml_tensor * sinks = dst->src[4];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+    if (sinks) {
+        GGML_ASSERT(sinks->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    cl_kernel kernel = NULL;
+
+    const bool is_f16 = q->type == GGML_TYPE_F16;
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16;
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+
+    if (n_q == 1) {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16_q1.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16_q1.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32_q1.at(dk_dv);
+        }
+    } else {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32.at(dk_dv);
+        }
+    }
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+    ggml_tensor_extra_cl * extra_sinks = sinks ? (ggml_tensor_extra_cl *)sinks->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+    cl_mem   sinks_buffer = extra_sinks ? extra_sinks->data_device : NULL;
+    cl_ulong offset_sinks = extra_sinks ? extra_sinks->offset + sinks->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
+    const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
+    const cl_ulong v_nb1 = v->nb[1], v_nb2 = v->nb[2], v_nb3 = v->nb[3];
+    const cl_ulong o_nb1 = dst->nb[1], o_nb2 = dst->nb[2], o_nb3 = dst->nb[3];
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float scale, max_bias, logit_softcap;
+    const float * params = (const float *)dst->op_params;
+    scale         = params[0];
+    max_bias      = params[1];
+    logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra_q->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset_q));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra_k->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset_k));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra_v->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset_v));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem),   &extra_o->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offset_o));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),     &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),     &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &q_nb1)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &q_nb2)); CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &k_nb1)); CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &k_nb2)); CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &v_nb1)); CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &v_nb2)); CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &o_nb1)); CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &o_nb2)); CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &offset_mask));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &mask_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &mask_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 38, sizeof(cl_mem),   &sinks_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 39, sizeof(cl_ulong), &offset_sinks));
+
+    if (n_q == 1) {
+        const size_t wg_size = 64;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    } else {
+        const int block_m = backend_ctx->kernels_flash_attn_bm.at(dk_dv);
+        const size_t wg_size = block_m;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { (size_t)((n_q + block_m - 1) / block_m) * wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int M = src0->ne[1];
+    const int N = src1->ne[1];
+    const int K = src0->ne[0];
+
+    cl_kernel kernel = backend_ctx->kernel_mul_mat_f16_f32_tiled;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(int),      &M));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(int),      &N));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),      &K));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &offsetd));
+
+    // Tiling parameters. These need to be tuned for optimal performance.
+    // They must match the #defines in the kernel mul_mat_f16_f32.cl.
+    //
+    // OPWM / OPWN: Output tile size per Work-Group. A work-group computes a tile of size OPWM x OPWN.
+    // TPWM / TPWN: Threads per Work-group. This is the work-group size.
+    // OPTM / OPTN: Output elements per Thread. Each thread computes OPTM x OPTN elements.
+    //
+    // The following relationships must hold:
+    //   OPWM = TPWM * OPTM
+    //   OPWN = TPWN * OPTN
+    //
+    const int OPWM = 64;
+    const int OPWN = 64;
+    const int TPWM = 16;
+    const int TPWN = 8;
+
+    size_t local_work_size[2] = { TPWM, TPWN };
+    size_t global_work_size[2] = {
+        (size_t) ((M + OPWM - 1) / OPWM) * TPWM,
+        (size_t) ((N + OPWN - 1) / OPWN) * TPWN,
+    };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_conv_2d(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const cl_uint Cout = ne03; const cl_uint Cin = ne02; const cl_uint N = ne13;
+    const cl_uint KW = ne00; const cl_uint KH = ne01; const cl_uint W = ne10; const cl_uint H = ne11; const cl_uint OW = ne0; const cl_uint OH = ne1;
+
+    const cl_uint s0 = dst->op_params[0]; const cl_uint s1 = dst->op_params[1];
+    const cl_uint p0 = dst->op_params[2]; const cl_uint p1 = dst->op_params[3];
+    const cl_uint d0 = dst->op_params[4]; const cl_uint d1 = dst->op_params[5];
+
+    const cl_uint cl_nb01 = nb01/ggml_type_size(src0->type); const cl_uint cl_nb02 = nb02/ggml_type_size(src0->type); const cl_uint cl_nb03 = nb03/ggml_type_size(src0->type);
+    const cl_uint cl_nb11 = nb11/ggml_type_size(src1->type); const cl_uint cl_nb12 = nb12/ggml_type_size(src1->type); const cl_uint cl_nb13 = nb13/ggml_type_size(src1->type);
+    const cl_uint cl_nb1 = nb1/ggml_type_size(dst->type); const cl_uint cl_nb2 = nb2/ggml_type_size(dst->type); const cl_uint cl_nb3 = nb3/ggml_type_size(dst->type);
+
+    const int64_t NPQ = (int64_t)N * OW * OH;
+
+    const uint32_t BS_K = 64;
+    const uint32_t BS_NPQ = 64;
+    const uint32_t BS_CRS = 16;
+    const uint32_t VEC_SIZE = 4;
+
+    const uint32_t TS_K = 4;
+    const uint32_t TS_NPQ = 8;
+
+    const uint32_t WG_K = BS_K / TS_K;
+    const uint32_t WG_NPQ = BS_NPQ / TS_NPQ;
+
+    auto splitWork = [](uint32_t work_size, uint32_t block_size) { return (block_size + work_size - 1) / block_size; };
+    const uint32_t NB_K = splitWork(Cout, BS_K);
+    const uint32_t NB_NPQ = splitWork(NPQ, BS_NPQ);
+
+    cl_kernel kernel;
+    size_t shmem_size;
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_conv_2d_f16;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_half4));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_float) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_conv_2d_f16_f32;
+        shmem_size = (size_t)(BS_K * BS_CRS * sizeof(cl_half) + BS_CRS * (BS_NPQ / VEC_SIZE) * sizeof(cl_float4));
+    } else {
+        GGML_ASSERT(false && "Unsupported data type combination for conv2d");
+    }
+
+    cl_uint idx = 0;
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra0->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extra1->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_mem), &extrad->data_device)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, idx++, shmem_size, NULL));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cout)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &Cin)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &N));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &KH)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &W)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &H));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OW)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &OH));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &s1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &p1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d0)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &d1));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb01)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb02)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb03));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb11)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb12)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb13));
+    CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb1)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb2)); CL_CHECK(clSetKernelArg(kernel, idx++, sizeof(cl_uint), &cl_nb3));
+
+    size_t global_work_size[] = { (size_t)NB_K * WG_K, (size_t)NB_NPQ * WG_NPQ, 1 };
+    size_t local_work_size[] = { (size_t)WG_K, (size_t)WG_NPQ, 1 };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_mul_mat_kq_kqv_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+    const int  ne02 = src0->ne[2];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+
+    const int  ne10 = src1->ne[0];
+    const int  ne11 = src1->ne[1];
+    const int  ne12 = src1->ne[2];
+
+    const cl_ulong nb10 = src1->nb[0];
+
+    const int  ne0 = dst->ne[0];
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 == ne10);
+
+    cl_kernel kernel;
+    cl_context context = backend_ctx->context;
+
+    cl_int              status;
+    cl_image_format     img_fmt_1d;
+    cl_image_desc       img_desc_1d;
+    cl_buffer_region    region;
+    cl_mem              A_image1d;
+    cl_mem              A_sub_buffer;
+    cl_mem              B_sub_buffer;
+    cl_mem              D_image1d;
+    cl_mem              D_sub_buffer;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    if (nb01 > nb02) {
+        // KQ
+        kernel = backend_ctx->kernel_mul_mm_f16_f32_kq;
+    } else {
+        // KQV
+        kernel = backend_ctx->kernel_mul_mm_f16_f32_kqv;
+    }
+    // create sub-buffer for A
+    // <--------------------------------------------> //
+    extra0 = src0->view_src ? (ggml_tensor_extra_cl *)src0->view_src->extra : (ggml_tensor_extra_cl *)src0->extra;
+
+    region.origin = (extra0->offset);
+    if (nb01 > nb02) {
+        // KQ
+        region.size = nb01 * ne01;
+    } else {
+        // KQV
+        region.size = nb02 * ne02;
+    }
+
+    A_sub_buffer = clCreateSubBuffer((extra0->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    // <--------------------------------------------> //
+
+    // create sub-buffer for B
+    // <--------------------------------------------> //
+    region.origin = (extra1->offset);
+    region.size = nb10 * ne10 * ne11 * ne12;
+    B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    img_fmt_1d = {CL_RGBA, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    if (nb01 > nb02) {
+        img_desc_1d.image_width = (nb01 * ne01 / 4)/4;
+    }
+    else {
+        img_desc_1d.image_width = (nb02 * ne02 / 4)/4;
+    }
+    img_desc_1d.buffer = A_sub_buffer;
+    A_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create sub-buffer for output C
+    // <--------------------------------------------> //
+    region.origin = (extrad->offset);
+    region.size = ne0 * ne1 * dst->ne[2] * dst->nb[0]; // size of C in bytes
+    D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    // create image for C output
+    // <--------------------------------------------> //
+    img_fmt_1d = {CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4;
+    img_desc_1d.buffer = D_sub_buffer;
+    D_image1d = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+    // <--------------------------------------------> //
+
+    int offset_src0 = 0;
+    int offset_src1 = 0;
+
+    // set kernel args
+    // <--------------------------------------------> //
+    cl_uint k_arg = 0;
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &A_image1d));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &offset_src0));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &B_sub_buffer));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &offset_src1));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem), &D_image1d));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &extrad->offset));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &M));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &K));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &N));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &ne12));
+    CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),    &nb01));
+
+    size_t global_work_size[3] = {64, static_cast<size_t>(((M+63)/64)), static_cast<size_t>(((N+31)/32)*ne12)};
+    size_t local_work_size[3] = {64, 1, 2};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+    // deallocate sub buffers and images
+    // <--------------------------------------------> //
+    CL_CHECK(clReleaseMemObject(A_image1d));
+    CL_CHECK(clReleaseMemObject(D_image1d));
+    CL_CHECK(clReleaseMemObject(A_sub_buffer));
+    CL_CHECK(clReleaseMemObject(B_sub_buffer));
+    CL_CHECK(clReleaseMemObject(D_sub_buffer));
+}
+
+static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const enum ggml_type src0t = src0->type;
+    const enum ggml_type src1t = src1->type;
+
+    GGML_ASSERT(src0t == GGML_TYPE_Q8_0);
+    GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+
+    GGML_ASSERT(src1->view_offs == 0);
+    GGML_ASSERT(dst->view_offs == 0);
+
+    const int  ne00 = src0->ne[0];
+    const int  ne01 = src0->ne[1];
+    const int  ne02 = src0->ne[2];
+
+    const int  ne10 = src1->ne[0];
+    const int  ne12 = src1->ne[2];
+
+    const int  ne0 = dst->ne[0];
+    const int  ne1 = dst->ne[1];
+
+    GGML_ASSERT(ne00 == ne10);
+    GGML_ASSERT((ne00 % 32) == 0);
+    GGML_ASSERT(ne0 == ne01);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    // init CL objects
+    cl_int              status;
+    cl_image_format     img_fmt_1d;
+    cl_image_desc       img_desc_1d;
+    cl_buffer_region    region;
+    cl_mem              A_image1d;
+    cl_mem              B_image1d;
+    cl_mem              B_sub_buffer;
+    cl_mem              S_image1d;
+
+    cl_mem              D_image1d;
+    cl_mem              D_sub_buffer;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    // create an image for A
+    img_fmt_1d = { CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * K / 4;    // Divide by 4 for char -> float
+    img_desc_1d.buffer = extra0_q8_0->q;
+    A_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create an image for Scale
+    img_fmt_1d = { CL_R, CL_HALF_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * K / 32;    // Block size is 32
+    img_desc_1d.buffer = extra0_q8_0->d;
+    S_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // create a sub_buffer for B
+    region.origin = (extra1->offset); // + src1->view_offs);
+    region.size = K * N * sizeof(float);
+    B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    // create an image for B from sub_buffer: RGBA (OCL)
+    img_fmt_1d = {CL_RGBA, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = K * N / 4;
+    img_desc_1d.buffer = B_sub_buffer;
+    B_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    // Create subbuffer and image1d_buffer for dst
+    region.origin = (extrad->offset); // + dst->view_offs;
+    region.size = M * N * sizeof(float);
+    D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+    CL_CHECK(status);
+
+    img_fmt_1d = {CL_R, CL_FLOAT};
+    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+    img_desc_1d.image_width = M * N;
+    img_desc_1d.buffer = D_sub_buffer;
+    D_image1d = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
+    CL_CHECK(status);
+
+    size_t local_work_size[3] = {1, 1, 1};
+    size_t global_work_size[3] = {1, 1, 1};
+
+    if (N == 1) {
+        kernel = backend_ctx->CL_mul_mat_vec_q8_0_f32;
+
+        int r2 = 1;
+        int r3 = 1;
+        cl_uint k_arg = 0;
+
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &A_image1d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extra0_q8_0->d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &B_image1d));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extra1->offset));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extrad->offset));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r2));
+        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r3));
+
+        size_t wavesize = backend_ctx->adreno_wave_size;
+        local_work_size[0] = wavesize;
+        local_work_size[1] = 4; // reduce factor
+        local_work_size[2] = 1;
+
+        global_work_size[0] = ((M + wavesize - 1) / wavesize) * wavesize;
+        global_work_size[1] = 4; // reduce factor
+        global_work_size[2] = 1;
+    } else {
+        cl_ulong offsetd = extrad->offset + dst->view_offs;
+        cl_mem              B_image1d_trans = nullptr;
+        // for B transpose
+        cl_mem B_d = nullptr;
+        int padding;
+
+        //how many extra elements beyond multiple of 8
+        int extra_elements = N % 8;
+
+        //how much padding to add
+        padding = 0;
+        if (extra_elements > 0){
+            padding = 8 - extra_elements;
+        }
+
+        // Specify the starting offset (in bytes)
+        region.origin = 0;
+        // Specify the size of the sub-buffer (divide by 2 for FP16)
+        region.size = K * (N + padding) * sizeof(float)/2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        B_d = clCreateSubBuffer(
+            backend_ctx->prealloc_act_trans.buffer,
+            0,
+            CL_BUFFER_CREATE_TYPE_REGION,
+            &region,
+            &status);
+        CL_CHECK(status);
+
+        cl_image_format image_format_B_d_output = { CL_RGBA, CL_HALF_FLOAT }; //(CL_HALF_FLOAT for FP16)
+        cl_image_desc image_desc_B_d_output = {
+            CL_MEM_OBJECT_IMAGE1D_BUFFER,
+            static_cast<size_t>(K * (N + padding)/4),
+            0, 0, 0, 0, 0, 0, 0, { B_d }
+        };
+        B_image1d_trans = clCreateImage(
+            context,
+            0,
+            &image_format_B_d_output,
+            &image_desc_B_d_output,
+            NULL,
+            &status);
+        CL_CHECK(status);
+
+        int height_B = N/4;
+        if (height_B == 0) {
+            height_B = 1;
+        }
+        int width_B = K/4;
+        int padded_height_B = (N + padding)/4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &B_image1d));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_size_t[2] = { 1, 16 };
+        size_t global_size_t[2] = {
+            static_cast<size_t>(width_B),
+            static_cast<size_t>(padded_height_B)
+        };
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
+
+        kernel = backend_ctx->kernel_mul_mm_q8_0_f32_8x4;
+
+        int N_with_padding = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &K));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &M));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &N_with_padding));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &N));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &offsetd));
+
+        global_work_size[0] = (size_t)(N + 7) / 8;
+        global_work_size[1] = (size_t)(M + 3) / 4;
+        global_work_size[2] = 1;
+
+        local_work_size[0] = 2;
+        local_work_size[1] = 128;
+        local_work_size[2] = 1;
+    }
+
+    // enqueue kernel with profiling
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+    // deallocate sub buffers and images
+    CL_CHECK(clReleaseMemObject(A_image1d));
+    CL_CHECK(clReleaseMemObject(B_sub_buffer));
+    CL_CHECK(clReleaseMemObject(B_image1d));
+    CL_CHECK(clReleaseMemObject(S_image1d));
+    CL_CHECK(clReleaseMemObject(D_sub_buffer));
+    CL_CHECK(clReleaseMemObject(D_image1d));
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
+static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+    const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+    ggml_tensor_extra_cl_q6_K * extra0_q6_K = (ggml_tensor_extra_cl_q6_K *)src0->extra;
+#endif
+
+    const int  ne00 = src0 ? src0->ne[0] : 0;
+    const int  ne01 = src0 ? src0->ne[1] : 0;
+    const int  ne02 = src0 ? src0->ne[2] : 0;
+    const int  ne03 = src0 ? src0->ne[3] : 0;
+
+    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
+    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
+    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
+    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+
+    const int  ne10 = src1 ? src1->ne[0] : 0;
+    const int  ne11 = src1 ? src1->ne[1] : 0;
+    const int  ne12 = src1 ? src1->ne[2] : 0;
+    const int  ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
+    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
+    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
+    const cl_ulong nb13 = src1 ? src1->nb[3] : 0;
+
+    const int  ne0 = dst ? dst->ne[0] : 0;
+    const int  ne1 = dst ? dst->ne[1] : 0;
+
+    int r2 = ne12/ne02;
+    int r3 = ne13/ne03;
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int nth0 = 32;
+    int nth1 = 1;
+    int nrows = 1;
+    // The number of values produced by each subgroup
+    int ndst = 4;
+
+    cl_kernel kernel;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    cl_context context = backend_ctx->context;
+
+    if(src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32){
+        if (ne01 >= 64 && ne1 >= 32 && ne00 >= 16 && (ne12 % ne02) == 0  &&
+            // dst is wrapped with image1d_buffer, the size limit applies, also src0
+            (ne0 * ne1 * dst->ne[2] * dst->nb[0] / 4 <= backend_ctx->image_max_buffer_size)) {
+            // For KQ
+            if (ggml_is_permuted(src0) && ggml_is_permuted(src1) &&
+                ((nb01 * ne01 / 4)/4 <= backend_ctx->image_max_buffer_size) &&
+                nb00 <= nb02 &&
+                nb02 <= nb01 &&
+                nb01 <= nb03 &&
+                nb10 <= nb12 &&
+                nb12 <= nb11 &&
+                nb11 <= nb13) {
+                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
+                return;
+            }
+            // For KQV
+            if (!ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+                ((nb02 * ne02 / 4)/4 <= backend_ctx->image_max_buffer_size)) {
+                ggml_cl_mul_mat_kq_kqv_adreno(backend, src0, src1, dst);
+                return;
+            }
+        }
+    }
+
+    if (ne01 && ne1 && use_adreno_kernels(backend_ctx, src0)) {
+
+    // init CL objects
+    // <--------------------------------------------> //
+    cl_int              status;
+    cl_image_format     img_fmt_1d;
+    cl_image_desc       img_desc_1d;
+    cl_buffer_region    region;
+    cl_mem              A_image1d = nullptr;
+    cl_mem              B_image1d = nullptr;
+    cl_mem              B_sub_buffer = nullptr;
+    cl_mem              C_d = nullptr;
+    // for B transpose
+    cl_mem B_d = nullptr;
+    cl_mem B_d_input_image = nullptr;
+    // <--------------------------------------------> //
+
+    // define matrix dimensions
+    // <--------------------------------------------> //
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+    int padding;
+    // <--------------------------------------------> //
+
+    // q8_0 x fp32
+    if (src0t == GGML_TYPE_Q8_0 && src1t == GGML_TYPE_F32 &&
+        enable_adreno_trans_weight(backend_ctx, src0)) {
+            ggml_cl_mul_mat_q8_0_f32_adreno(backend, src0, src1, dst);
+            return;
+    }
+
+    // q4_0 x fp32
+    if(src0t == GGML_TYPE_Q4_0 && src1t == GGML_TYPE_F32) {
+        // TODO: remove duplicate definitions of image description + format -- move to top
+
+        // create an image for A
+        // <--------------------------------------------> //
+        if (N == 1) {
+            img_fmt_1d = { CL_R, CL_UNSIGNED_INT32};
+        } else {
+            img_fmt_1d = { CL_R, CL_FLOAT};
+        }
+        memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc_1d.image_width = M * K / 2 / 4;    // Divide by 4 for char -> float
+        img_desc_1d.buffer = extra0_q4_0->q;
+        A_image1d = clCreateImage(
+            context,
+            CL_MEM_READ_ONLY,
+            &img_fmt_1d,
+            &img_desc_1d,
+            NULL,
+            &status);
+        CL_CHECK(status);
+        // <--------------------------------------------> //
+
+
+        // create a sub_buffer for B
+        // <--------------------------------------------> //
+        region.origin = (extra1->offset);
+        region.size = K * N * sizeof(float);
+        B_sub_buffer = clCreateSubBuffer(
+            extra1->data_device,
+            0,
+            CL_BUFFER_CREATE_TYPE_REGION,
+            &region,
+            &status);
+        CL_CHECK(status);
+        // <--------------------------------------------> //
+
+        // transpose activation for Skyler's gemm
+        if (N != 1) {
+            //how many extra elements beyond multiple of 8
+            int extra_elements = N % 8;
+
+            //how much padding to add
+            padding = 0;
+            if (extra_elements > 0){
+                padding = 8 - extra_elements;
+            }
+
+            // Specify the starting offset (in bytes)
+            region.origin = 0;
+            // Specify the size of the sub-buffer (divide by 2 for FP16)
+            region.size = K * (N + padding) * sizeof(float)/2;
+            backend_ctx->prealloc_act_trans.allocate(context, region.size);
+
+            B_d = clCreateSubBuffer(
+                backend_ctx->prealloc_act_trans.buffer,
+                0,
+                CL_BUFFER_CREATE_TYPE_REGION,
+                &region,
+                &status);
+            CL_CHECK(status);
+
+            cl_image_format image_format_B_d_input = { CL_RGBA, CL_FLOAT };
+            cl_image_desc image_desc_B_d_input = {
+                CL_MEM_OBJECT_IMAGE1D_BUFFER,
+                static_cast<size_t>(K * N / 4),
+                0, 0, 0, 0, 0, 0, 0, { B_sub_buffer }
+            };
+            B_d_input_image = clCreateImage(
+                context,
+                0,
+                &image_format_B_d_input,
+                &image_desc_B_d_input,
+                NULL,
+                &status);
+            CL_CHECK(status);
+
+            cl_image_format image_format_B_d_output = { CL_RGBA, CL_HALF_FLOAT }; //(CL_HALF_FLOAT for FP16)
+            cl_image_desc image_desc_B_d_output = {
+                CL_MEM_OBJECT_IMAGE1D_BUFFER,
+                static_cast<size_t>(K * (N + padding)/4),
+                0, 0, 0, 0, 0, 0, 0, { B_d }
+            };
+            B_image1d = clCreateImage(
+                context,
+                0,
+                &image_format_B_d_output,
+                &image_desc_B_d_output,
+                NULL,
+                &status);
+            CL_CHECK(status);
+
+            int height_B = N/4;
+            if (height_B == 0) {
+                height_B = 1;
+            }
+            int width_B = K/4;
+            int padded_height_B = (N + padding)/4;
+
+            kernel = backend_ctx->kernel_transpose_32_16;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &B_d_input_image));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &B_image1d));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+            size_t local_size_t[2] = { 1, 16 };
+            //WGS tuning
+            if (ne0 == 4096 && ne1 == 128 && ne10 == 4096) {
+                local_size_t[0]=4;
+                local_size_t[1]=8;
+            } else if (ne0 == 11008 && ne1 == 128 && ne10 == 4096) {
+                local_size_t[0]=2;
+                local_size_t[1]=8;
+            } else if(ne0 == 4096 && ne1 == 128 && ne10 == 11008) {
+                local_size_t[0]=1;
+                local_size_t[1]=8;
+            } else if(ne0 == 32000 && ne1 == 128 && ne10 == 4096) {
+                local_size_t[0]=2;
+                local_size_t[1]=8;
+            }
+
+            size_t global_size_t[2] = {
+                static_cast<size_t>(width_B),
+                static_cast<size_t>(padded_height_B)
+            };
+
+            backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
+        } else {
+            // no need to transpose B in other cases
+            // create an image for B from sub_buffer
+            // <--------------------------------------------> //
+            img_fmt_1d = {CL_RGBA, CL_FLOAT};
+
+            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+            img_desc_1d.image_width = K * N / 4;
+            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+            img_desc_1d.buffer = B_sub_buffer;
+            B_image1d = clCreateImage(
+                context,
+                CL_MEM_READ_ONLY,
+                &img_fmt_1d,
+                &img_desc_1d,
+                NULL,
+                &status);
+            CL_CHECK(status);
+            // <--------------------------------------------> //
+        }
+
+        // choose gemm or gemv kernel
+        // <--------------------------------------------> //
+        if (N == 1) {
+            kernel = backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_general;
+            if (M == 4096 && K == 4096) {
+                kernel = backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_4096;
+            } else if (M == 4096 && K == 11008) {
+                kernel = backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_4096_1_11008;
+            } else if (M == 11008 && K == 4096) {
+                kernel = backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_11008_1_4096;
+            } else if (M == 32000 && K == 4096) {
+                kernel = backend_ctx->CL_mul_mat_vec_q4_0_f32_1d_4x_flat_32000_1_4096;
+            }
+        } else {
+            kernel = backend_ctx->CL_mul_mat_Ab_Bi_8x4;
+        }
+        // <--------------------------------------------> //
+
+        // set kernel args
+        // <--------------------------------------------> //
+        cl_uint k_arg = 0;
+
+        if (N == 1) {
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &A_image1d));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &B_image1d));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extra1->offset));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extrad->offset));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r3));
+        } else {
+            region.origin = extrad->offset; // Specify the starting offset (in bytes)
+            region.size = M * N * sizeof(float); // Specify the size of the sub-buffer
+            C_d = clCreateSubBuffer(extrad->data_device, CL_MEM_WRITE_ONLY, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+            CL_CHECK(status);
+
+            int padded_N = ne1 + padding;
+
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &extra0_q4_0->q)); //A_q_dextra0_q4_0->q
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &extra0_q4_0->d)); //A_s_d
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem), &B_image1d)); //B_d
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem), &C_d)); //C_d
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &ne01)); //M
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),    &padded_N)); //N with padding
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),    &ne00)); //K
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),    &ne1)); //N without padding
+        }
+        // <--------------------------------------------> //
+
+        // choose workgroup size
+        // <--------------------------------------------> //
+        size_t global_work_size[3] = {
+            64, static_cast<size_t>((M+63)/64), static_cast<size_t>((N+31)/32)};
+        size_t local_work_size[3] = {64, 2, 4};
+
+        global_work_size[0] = (size_t)(ceil((float)ne1/8));
+        global_work_size[1] = (size_t)(ne01/4);
+        global_work_size[2] = (size_t)(1);
+
+        local_work_size[0]  = (size_t)(1); //4x32 for FP32
+        local_work_size[1]  = (size_t)(128);
+        local_work_size[2]  = (size_t)(1);
+
+        //WGS tuning
+        if (ne0 == 4096 && ne1 == 128 && ne10 == 4096) {
+            local_work_size[0] = 1;
+            local_work_size[1] = 128;
+        } else if (ne0 == 11008 && ne1 == 128 && ne10 == 4096) {
+            local_work_size[0] = 2;
+            local_work_size[1] = 64;
+        } else if (ne0 == 4096 && ne1 == 128 && ne10 == 11008) {
+            local_work_size[0] = 2;
+            local_work_size[1] = 64;
+        } else if (ne0 == 32000 && ne1 == 128 && ne10 == 4096) {
+            local_work_size[0] = 2;
+            local_work_size[1] = 64;
+        }
+
+        if (N == 1) {
+            size_t wavesize = backend_ctx->adreno_wave_size;
+            local_work_size[0] = wavesize; // localsize
+            local_work_size[1] = 4; // reduce factor
+            local_work_size[2] = 1;
+
+            global_work_size[0] = (((M / 2) + wavesize - 1) / wavesize) * wavesize;
+            global_work_size[1] = 4; // reduce factor
+            global_work_size[2] = 1;
+        }
+        // <--------------------------------------------> //
+
+        // enqueue kernel with profiling
+        // <--------------------------------------------> //
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+        // <--------------------------------------------> //
+
+        // deallocate sub buffers and images
+        // <--------------------------------------------> //
+        CL_CHECK(clReleaseMemObject(A_image1d));
+        CL_CHECK(clReleaseMemObject(B_sub_buffer));
+        CL_CHECK(clReleaseMemObject(B_image1d));
+
+        if (N != 1) {
+            CL_CHECK(clReleaseMemObject(B_d));
+            CL_CHECK(clReleaseMemObject(B_d_input_image));
+            CL_CHECK(clReleaseMemObject(C_d));
+        }
+        // <--------------------------------------------> //
+
+        return;
+    }
+    } // if (ne01 && ne1)
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+    // GEMM using local memory
+    // Current BK = 16, so ne00 % 16 == 0
+    if (src1t == GGML_TYPE_F32 &&
+        ne00 % 16 == 0 &&
+        ne11 > 1) {
+        switch(src0t) {
+            case GGML_TYPE_F32: {
+                kernel = backend_ctx->kernel_mul_mm_f32_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                        nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_F16: {
+                kernel = backend_ctx->kernel_mul_mm_f16_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                cl_mem mem_src0 = extra0->data_device;
+                cl_mem mem_src1 = extra1->data_device;
+
+                cl_ulong nb00_cont = nb00;
+                cl_ulong nb01_cont = nb01;
+                cl_ulong nb02_cont = nb02;
+                cl_ulong nb03_cont = nb03;
+
+                cl_ulong nb10_cont = nb10;
+                cl_ulong nb11_cont = nb11;
+                cl_ulong nb12_cont = nb12;
+                cl_ulong nb13_cont = nb13;
+
+                cl_ulong offset0_cont = offset0;
+                cl_ulong offset1_cont = offset1;
+
+                if (!ggml_is_contiguous(src0)) {
+                    backend_ctx->prealloc_src0.allocate(backend_ctx->context, ggml_nbytes(src0));
+                    ggml_cl_copy_to_contiguous(backend, src0, backend_ctx->prealloc_src0.buffer,
+                        nb00_cont, nb01_cont, nb02_cont, nb03_cont);
+                    mem_src0 = backend_ctx->prealloc_src0.buffer;
+                    offset0_cont = 0;
+                }
+
+                if (!ggml_is_contiguous(src1)) {
+                    backend_ctx->prealloc_src1.allocate(backend_ctx->context, ggml_nbytes(src1));
+                    ggml_cl_copy_to_contiguous(backend, src1, backend_ctx->prealloc_src1.buffer,
+                            nb10_cont, nb11_cont, nb12_cont, nb13_cont);
+                    mem_src1 = backend_ctx->prealloc_src1.buffer;
+                    offset1_cont = 0;
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &mem_src0));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0_cont));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &mem_src1));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1_cont));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_Q8_0: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q8_0_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            case GGML_TYPE_Q6_K: {
+                if (ne11 < 32) {
+                    break;
+                }
+                if (!ggml_is_contiguous(src0) || !ggml_is_contiguous(src1)) {
+                    break;
+                }
+
+                kernel = backend_ctx->kernel_mul_mm_q6_k_f32_l4_lm;
+                nth0 = 128; // calculated as (BM*BN)/(TM*TN)
+
+                int batch_stride_a = ne00*ne01;
+                int batch_stride_b = ne10*ne11;
+                int batch_stride_d = ne0*ne1;
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q6_K->qh));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q6_K->s));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q6_K->d));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10)); // stride_a
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne10)); // stride_b
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne01)); // stride_d
+                CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &batch_stride_a));
+                CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &batch_stride_b));
+                CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &batch_stride_d));
+                CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &r3));
+
+                // 64 is block tile size BM and BN - change here when BM and BN in the kernel are changed.
+                size_t global_work_size[] = {(size_t)(CEIL_DIV(ne01, 64)*nth0), (size_t)(CEIL_DIV(ne11, 64)), (size_t)ne12*ne13};
+                size_t local_work_size[] = {(size_t)nth0, 1, 1};
+
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+                return;
+            }
+            default:
+                break;
+        }
+    }
+
+    if (src0t == GGML_TYPE_F16 && src1t == GGML_TYPE_F32 &&
+        src0->ne[1] > 32 &&   // M > 32
+        src1->ne[1] > 32 &&   // N > 32
+        src0->ne[0] > 32 &&   // K > 32
+        src0->ne[2] == 1 && src0->ne[3] == 1 &&
+        src1->ne[2] == 1 && src1->ne[3] == 1 &&
+        ggml_is_contiguous(src0) && ggml_is_contiguous(src1) &&
+        backend_ctx->kernel_mul_mat_f16_f32_tiled != NULL) {
+        ggml_cl_mul_mat_f16_f32_tiled(backend, src0, src1, dst);
+        return;
+    }
+
+    if (!ggml_is_transposed(src0) &&
+        !ggml_is_transposed(src1) &&
+        src1t == GGML_TYPE_F32 &&
+        ne00%32 == 0 &&
+        ne11 > 2) {
+#ifdef GGML_OPENCL_SOA_Q
+        // Set up kernel.
+        switch(src0t) {
+            case GGML_TYPE_Q4_0:
+                // This should have been satisfied.
+                GGML_ASSERT(ne11 == ne1);
+                GGML_ASSERT(ne01 == ne0);
+
+                if (backend_ctx->gpu_family == INTEL) {
+                    nth0 = 16;
+                    nth1 = 1;
+
+                    kernel = backend_ctx->kernel_mul_mat_q4_0_f32_1d_16x_flat;
+                } else if (backend_ctx->gpu_family == ADRENO) {
+                    nth0 = 64;
+                    nth1 = 1;
+
+                    kernel = backend_ctx->kernel_mul_mat_q4_0_f32_1d_8x_flat;
+                } else {
+                    GGML_ASSERT(false && "TODO: Unknown GPU");
+                }
+
+                CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+                CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+                CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+                CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+                CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+                CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+                CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+                CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+                break;
+            default:
+                break;
+        }
+
+        // Launch kernel.
+        if (src0t == GGML_TYPE_Q4_0) {
+            size_t global_work_size[] = {(size_t)(ne01 + 7)/8*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+            size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+            if (backend_ctx->gpu_family == INTEL) {
+                // Set global size for Intel. It uses 16x output values.
+                global_work_size[0] = (size_t)(ne01 + 15)/16*nth0;
+                global_work_size[1] = (size_t)ne11*nth1;
+                global_work_size[2] = (size_t)ne12*ne13;
+            }
+
+            backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+            return;
+        }
+#else // GGML_OPENCL_SOA_Q
+        // TODO: add block_q4_0 variant.
+#endif // GGML_OPENCL_SOA_Q
+    }
+
+    // use custom matrix x vector kernel
+    switch (src0t) {
+        case GGML_TYPE_F32:
+            //GGML_ASSERT(ne02 == ne12);
+            GGML_ASSERT(src1t == GGML_TYPE_F32);
+            kernel = backend_ctx->kernel_mul_mat_f32_f32;
+            nrows = 4;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 32;
+                nth1 = 1;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            break;
+        case GGML_TYPE_F16:
+            //GGML_ASSERT(ne02 == ne12);
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 32;
+                nth1 = 1;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            if (src1t == GGML_TYPE_F32) {
+                if (ne11 * ne12 < 4) {
+                    kernel = backend_ctx->kernel_mul_mat_f16_f32_1row;
+                } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                    kernel = backend_ctx->kernel_mul_mat_f16_f32_l4;
+                    nrows = ne11;
+                } else {
+                    kernel = backend_ctx->kernel_mul_mat_f16_f32;
+                    nrows = 4;
+                }
+            } else {
+                kernel = backend_ctx->kernel_mul_mat_f16_f16;
+                nrows = 4;
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb10));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            break;
+        case GGML_TYPE_Q4_0:
+            // This should have been satisfied.
+            GGML_ASSERT(ne11 == ne1);
+            GGML_ASSERT(ne01 == ne0);
+
+#ifdef GGML_OPENCL_SOA_Q
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+
+                kernel = backend_ctx->kernel_mul_mat_q4_0_f32_8x_flat;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+
+                kernel = backend_ctx->kernel_mul_mat_q4_0_f32_8x_flat;
+                ndst =8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#else // GGML_OPENCL_SOA_Q
+            if (backend_ctx->gpu_family == INTEL) {
+                // Use 1D local size. Each workgroup is a SIMD group. Each SIMD
+                // group produces N_DST (4 for Q4_0 kernel) values in the result.
+                // The number of workgroups on dim 0 (the leading dimension) is
+                // the nearest multiple of 4 that covers ne0 (equals ne01).
+                nth0 = 16;
+                nth1 = 1;
+
+                kernel = backend_ctx->kernel_mul_mat_q4_0_f32;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+
+                kernel = backend_ctx->kernel_mul_mat_q4_0_f32_v;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32_flat;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_q8_0_f32;
+
+            // nth0 - subgroup size
+            // nth1 - number of subgroups per workgroup
+            // ndst - number of output values per workgroup = output per subgroup * number of subgroups
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K: {
+            kernel = backend_ctx->kernel_mul_mv_q4_K_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 1;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 1;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),     &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(int),        &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),     &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(int),        &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),     &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),        &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),        &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),        &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong),   &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong),   &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),   &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),        &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),   &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),   &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),   &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),        &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),        &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),        &r2));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),        &r3));
+            break;
+        }
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_q6_K_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q6_K->ql));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q6_K->qh));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q6_K->s));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q6_K->d));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_q6_K_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = 1;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = 1;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+#else
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+#endif
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
+        src0t == GGML_TYPE_Q4_1 ||
+        src0t == GGML_TYPE_Q8_0 ||
+        src0t == GGML_TYPE_Q2_K) {
+        // Each SIMD group produces N_DST values in the result. Assuming each
+        // workgroup has N_SIMDGROUP SIMD groups, then each workgroup will
+        // produce N_DST*N_SIMDGROUP values in the result. Hence, the grid size
+        // (number of workgroups) will be a nearest multiple of
+        // N_DST*N_SIMDGROUP to cover the size of the dimension. Below, 4 is
+        // N_DST*N_SIMDGROUP (see the kernel for Q4_0 matmul).
+        size_t global_work_size[] = {(size_t)(ne01 + ndst-1)/ndst*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else if (src0t == GGML_TYPE_Q4_K) {
+        size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else if (src0t == GGML_TYPE_Q3_K) {
+        GGML_ASSERT(false && "not implemented");
+    } else if (src0t == GGML_TYPE_Q5_K) {
+        GGML_ASSERT(false && "not implemented");
+    } else if (src0t == GGML_TYPE_Q6_K) {
+        size_t global_work_size[] = {(size_t)(ne01+ndst*nth1-1)/(ndst*nth1)*nth0, (size_t)ne11*nth1, (size_t)ne12*ne13};
+        size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        int64_t ny = (ne11 + nrows - 1)/nrows;
+
+        size_t global_work_size[] = {(size_t)ne01*nth0, (size_t)ny*nth1, (size_t)ne12*ne13};
+        size_t local_work_size[] = {(size_t)nth0, (size_t)nth1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    }
+}
+
+static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    const ggml_tensor * src2 = dst->src[2];
+    GGML_ASSERT(src2);
+    GGML_ASSERT(src2->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offset2 = extra2->offset + src2->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    GGML_UNUSED(offset0);
+
+#ifdef GGML_OPENCL_SOA_Q
+    ggml_tensor_extra_cl_q4_0 * extra0_q4_0 = (ggml_tensor_extra_cl_q4_0 *)src0->extra;
+    ggml_tensor_extra_cl_mxfp4 * extra0_mxfp4 = (ggml_tensor_extra_cl_mxfp4 *)src0->extra;
+    ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
+#endif
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = src1->ne[0];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];
+
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const int ne20 = src2->ne[0];
+    const int ne21 = src2->ne[1];
+
+    const cl_ulong nb21 = src2->nb[1];
+    const cl_ulong nb20 = src2->nb[0];
+
+    UNUSED(nb20);
+
+    const int ne0 = dst->ne[0];
+    const int ne1 = dst->ne[1];
+
+    const int r2 = ne12/ne02;
+    const int r3 = ne13/ne03;
+    const int dst_rows = ne20*ne21; // ne20 = n_used_experts, ne21 = n_rows
+
+    GGML_ASSERT(ne00 == ne10);
+
+    int sgs   = 32; // subgroup size
+    int nsg   = 1;  // number of subgroups
+    int nrows = 1;  // number of row in src1
+    int ndst  = 4;  // number of values produced by each subgroup
+
+    cl_kernel kernel;
+
+    // subgroup mat vec
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0: {
+            kernel = backend_ctx->kernel_mul_mv_id_q4_0_f32_8x_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 1;
+                ndst = 8;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 8;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q4_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q4_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb00));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne10));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &r3));
+
+            break;
+        }
+        case GGML_TYPE_Q8_0: {
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32_flat;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#else
+            kernel = backend_ctx->kernel_mul_mv_id_q8_0_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 4;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 4;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne1));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
+        case GGML_TYPE_MXFP4: {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+            if (use_adreno_moe_kernels(backend_ctx, src0)) {
+                cl_int status;
+
+                size_t local_size[3] = {64, 2, 1};
+                size_t global_size[3] = {64, 2, 1};
+
+                cl_mem src1_sub_buffer, buf_src1_image, buf_src2;
+
+                int tile_size = 320;
+                if (ne12 == 1) { // for gemv
+                    kernel = backend_ctx->kernel_gemv_moe_mxfp4_f32;
+
+                    // create a sub_buffer for src2
+                    cl_buffer_region region;
+                    region.origin = offset2;
+                    region.size = ne20 * ne21 * sizeof(int);
+                    buf_src2 = clCreateSubBuffer(extra2->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+                    CL_CHECK(status);
+
+                    // set thread grid
+                    global_size[0] = static_cast<size_t>(ne01);
+                    global_size[1] = 4;
+                    global_size[2] = static_cast<size_t>(ne20);
+                    local_size[1] = 4;
+                } else { // for gemm
+                    kernel = backend_ctx->kernel_gemm_moe_mxfp4_f32;
+
+                    // preprocess router table
+                    int num_tiles_per_expert = (ne01 + tile_size - 1) / tile_size;
+                    void * host_src2_reorder = malloc(ne20 * ne21 * 4 * num_tiles_per_expert * sizeof(short));
+                    void * host_src2 = malloc(ne21 * nb21);
+                    CL_CHECK(clEnqueueReadBuffer(backend_ctx->queue, extra2->data_device, CL_TRUE, offset2, ne21 * nb21, host_src2, 0, NULL, NULL));
+                    int total_experts = nb21 / nb20;
+                    int out_idx = 0;
+                    for (int i_expert = 0; i_expert < ne02; i_expert++) {
+                        for (int i_tile = 0; i_tile < num_tiles_per_expert; i_tile++) {
+                            for (int j = 0; j < ne21; j++) {
+                                for (int i = 0; i < ne20; i++) {
+                                    int expert = ((int *)host_src2)[j * total_experts + i];
+                                    if (i_expert == expert) {
+                                        ((short *)host_src2_reorder)[out_idx] = static_cast<short>(expert);
+                                        ((short *)host_src2_reorder)[out_idx + 1] = static_cast<short>(j * ne11 + (i % ne11));
+                                        ((short *)host_src2_reorder)[out_idx + 2] = static_cast<short>(j * ne20 + i);
+                                        ((short *)host_src2_reorder)[out_idx + 3] = static_cast<short>(i_tile);
+                                        out_idx += 4;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                    buf_src2 = clCreateBuffer(backend_ctx->context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, ne20 * ne21 * 4 * num_tiles_per_expert * sizeof(short), host_src2_reorder, &status);
+                    CL_CHECK(status);
+
+                    // set thread grid
+                    global_size[0] = static_cast<size_t>(tile_size);
+                    global_size[2] = static_cast<size_t>(ne20 * ne21 * num_tiles_per_expert);
+                }
+
+                // create a sub_buffer for src1
+                cl_buffer_region region;
+                region.origin = offset1;
+                region.size = ne10 * ne11 * ne12 * sizeof(float);
+                src1_sub_buffer = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
+                CL_CHECK(status);
+
+                // create image for src1
+                cl_image_format image_format_buf_src1 = {CL_RGBA, CL_FLOAT};
+                cl_image_desc image_desc_buf_src1 = {CL_MEM_OBJECT_IMAGE1D_BUFFER, static_cast<size_t>(ne10 * ne11 * ne12 / 4), 0,0,0,0,0,0,0, {src1_sub_buffer}};
+                buf_src1_image = clCreateImage(backend_ctx->context, CL_MEM_READ_ONLY, &image_format_buf_src1, &image_desc_buf_src1, NULL, &status);
+                CL_CHECK(status);
+
+                // Set kernel args
+                int arg_idx = 0;
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &extra0_mxfp4->q));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &extra0_mxfp4->e));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &buf_src1_image));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &buf_src2));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_mem),    &extrad->data_device));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(cl_ulong),  &offsetd));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne00));
+                CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne01));
+                if (ne12 == 1) {
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &ne11));
+                } else {
+                    CL_CHECK(clSetKernelArg(kernel, arg_idx++, sizeof(int),       &tile_size));
+                }
+
+                // launch kernel
+                backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_size, local_size, dst);
+
+                // deallocate sub buffers and images
+                CL_CHECK(clReleaseMemObject(src1_sub_buffer));
+                CL_CHECK(clReleaseMemObject(buf_src1_image));
+                CL_CHECK(clReleaseMemObject(buf_src2));
+                return;
+            } // else fallback to generic kernel
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
+#ifdef GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32_flat;
+
+            cl_mem q;
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+
+                q = extra0_mxfp4->q;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 1;
+                ndst = 4;
+
+                q = extra0_mxfp4->q_img;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_mxfp4->e));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+#else // GGML_OPENCL_SOA_Q
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
+#endif // GGML_OPENCL_SOA_Q
+            break;
+        }
+        default:
+            GGML_ASSERT(false && "not implemented");;
+    }
+
+    int _ne1 = 1;
+    int ne123 = dst_rows;
+
+    size_t global_work_size[] = {(size_t)(ne01+ndst*nsg-1)/(ndst*nsg)*sgs, (size_t)(_ne1+nrows-1)/nrows*nsg, (size_t)ne123};
+    size_t local_work_size[] = {(size_t)sgs, (size_t)nsg, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_scale(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    float scale;
+    float bias;
+    memcpy(&scale, ((int32_t *) dst->op_params) + 0, sizeof(float));
+    memcpy(&bias,  ((int32_t *) dst->op_params) + 1, sizeof(float));
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    int n = ggml_nelements(dst);
+
+    if (n % 4 == 0) {
+        kernel = backend_ctx->kernel_scale_f32_4;
+        n /= 4;
+    } else {
+        kernel = backend_ctx->kernel_scale_f32;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(float),    &bias));
+
+    size_t global_work_size[] = {(size_t)n, 1, 1};
+    size_t local_work_size[] = {64, 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (n % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+    }
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_cpy(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+
+    // GGML_OP_CPY happens between src0 and src1.
+    // GGML_OP_DUP and GGML_OP_CONT happen between src0 and dst.
+    UNUSED(dst);
+
+    const int ne00 = src0 ? src0->ne[0] : 0;
+    const int ne01 = src0 ? src0->ne[1] : 0;
+    const int ne02 = src0 ? src0->ne[2] : 0;
+    const int ne03 = src0 ? src0->ne[3] : 0;
+
+    const cl_ulong nb00 = src0 ? src0->nb[0] : 0;
+    const cl_ulong nb01 = src0 ? src0->nb[1] : 0;
+    const cl_ulong nb02 = src0 ? src0->nb[2] : 0;
+    const cl_ulong nb03 = src0 ? src0->nb[3] : 0;
+
+    const int ne10 = src1 ? src1->ne[0] : 0;
+    const int ne11 = src1 ? src1->ne[1] : 0;
+    const int ne12 = src1 ? src1->ne[2] : 0;
+    const int ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_ulong nb10 = src1 ? src1->nb[0] : 0;
+    const cl_ulong nb11 = src1 ? src1->nb[1] : 0;
+    const cl_ulong nb12 = src1 ? src1->nb[2] : 0;
+    const cl_ulong nb13 = src1 ? src1->nb[3] : 0;
+
+    const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+    const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+
+    cl_kernel kernel;
+
+    switch (src0t) {
+        case GGML_TYPE_F32:
+            switch (src1t) {
+                case GGML_TYPE_F16:
+                    kernel = backend_ctx->kernel_cpy_f32_f16;
+                    break;
+                case GGML_TYPE_F32:
+                    kernel = backend_ctx->kernel_cpy_f32_f32;
+                    break;
+                default:
+                    GGML_ASSERT(false && "not implemented");
+            }
+            break;
+        case GGML_TYPE_F16:
+            switch (src1t) {
+                case GGML_TYPE_F16:
+                    kernel = backend_ctx->kernel_cpy_f16_f16;
+                    break;
+                case GGML_TYPE_F32:
+                    kernel = backend_ctx->kernel_cpy_f16_f32;
+                    break;
+                default:
+                    GGML_ASSERT(false && "not implemented");
+            }
+            break;
+        default:
+            GGML_ASSERT(false && "not implemented");
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb13));
+
+    const int nth = MIN(64, ne00);
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, src1);
+}
+
+static void ggml_cl_dup(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cl_cpy(backend, src0, dst, nullptr);
+    UNUSED(src1);
+}
+
+static void ggml_cl_diag_mask_inf(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    int n_past = ((int32_t *)(dst->op_params))[0];
+
+    const int  ne00 = src0 ? src0->ne[0] : 0;
+    const int  ne01 = src0 ? src0->ne[1] : 0;
+    const int  ne02 = src0 ? src0->ne[2] : 0;
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+
+    if (ne00%8 == 0) {
+        kernel = backend_ctx->kernel_diag_mask_inf_8;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &n_past));
+
+        size_t global_work_size[] = {(size_t)ne00*ne01*ne02/8, 1, 1};
+        size_t local_work_size[] = {64, 1, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+    } else {
+        kernel = backend_ctx->kernel_diag_mask_inf;
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &n_past));
+
+        size_t global_work_size[] = {(size_t)ne00, (size_t)ne01, (size_t)ne02};
+        size_t local_work_size[] = {64, 1, 1};
+
+        size_t * local_work_size_ptr = local_work_size;
+        if (ne00 % 64 != 0 && !backend_ctx->non_uniform_workgroups) {
+            local_work_size_ptr = nullptr;  // Let driver choose the work-group sizes.
+        }
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+    }
+}
+
+static void ggml_cl_soft_max(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // Softmax can now fuse KQ mask and KQ scale, which used to be two additional
+    // ops before softmax. It now also fuses alibi if `max_bias > 0`. For llama,
+    // alibi is not used; however, for some other models, it is used.
+    // KQ_mask
+    if (src1) {
+        GGML_ASSERT(src1);
+        GGML_ASSERT(src1->extra);
+    }
+
+    const ggml_tensor * src2 = dst->src[2];
+    if (src2) {
+        GGML_ASSERT(src2->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
+    ggml_tensor_extra_cl * extra2 = src2 ? (ggml_tensor_extra_cl *)src2->extra : nullptr;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
+    cl_ulong offset2 = extra2 ? extra2->offset + src2->view_offs : offset0;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_long nb01 = src0->nb[1];
+    const cl_long nb02 = src0->nb[2];
+    const cl_long nb03 = src0->nb[3];
+
+    const int ne12 = src1 ? src1->ne[2] : 0;
+    const int ne13 = src1 ? src1->ne[3] : 0;
+
+    const cl_long nb11 = src1 ? src1->nb[1] : 0;
+    const cl_long nb12 = src1 ? src1->nb[2] : 0;
+    const cl_long nb13 = src1 ? src1->nb[3] : 0;
+
+    const cl_long nb1 = dst->nb[1];
+    const cl_long nb2 = dst->nb[2];
+    const cl_long nb3 = dst->nb[3];
+
+    float scale, max_bias;
+    memcpy(&scale,    dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, dst->op_params + 1, sizeof(float));
+
+    const int n_head      = src0->ne[2];
+    const int n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
+    // Local size must be wave size. Each workgroup is a wave, working on a row,
+    // where a row corresponds to leading dimension.
+    int nth = MIN(32, ne00);
+
+    if (backend_ctx->gpu_family == INTEL) {
+        // This is the same as the initial value.
+        nth = MIN(32, ne00);
+    }
+    else if (backend_ctx->gpu_family == ADRENO) {
+        nth = 64;
+    } else {
+        GGML_ASSERT(false && "TODO: Unknown GPU");
+    }
+
+    cl_kernel kernel;
+
+    if (ne00%4 == 0) {
+        if (use_f16) {
+            kernel = backend_ctx->kernel_soft_max_4_f16;
+        } else {
+            kernel = backend_ctx->kernel_soft_max_4;
+        }
+    } else {
+        if (use_f16) {
+            kernel = backend_ctx->kernel_soft_max_f16;
+        } else {
+            kernel = backend_ctx->kernel_soft_max;
+        }
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   extra1 ? &extra1->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   extra2 ? &extra2->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &n_head_log2));
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_rope(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    ggml_tensor * src2 = dst->src[2];
+    ggml_tensor_extra_cl * extra2 = src2 ? (ggml_tensor_extra_cl *)src2->extra : nullptr;
+
+    cl_ulong offset2 = extra2 ? extra2->offset + src2->view_offs : offset0;
+
+    const int  ne00 = src0 ? src0->ne[0] : 0;
+    const int  ne01 = src0 ? src0->ne[1] : 0;
+    const int  ne02 = src0 ? src0->ne[2] : 0;
+    const int  ne03 = src0 ? src0->ne[3] : 0;
+
+    const cl_ulong  nb00 = src0 ? src0->nb[0] : 0;
+    const cl_ulong  nb01 = src0 ? src0->nb[1] : 0;
+    const cl_ulong  nb02 = src0 ? src0->nb[2] : 0;
+    const cl_ulong  nb03 = src0 ? src0->nb[3] : 0;
+
+    const int ne10 = src1 ? src1->ne[0] : 0;
+    const int ne11 = src1 ? src1->ne[1] : 0; UNUSED(ne11);
+    const int ne12 = src1 ? src1->ne[2] : 0; UNUSED(ne12);
+    const int ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+
+    const int  ne0 = dst ? dst->ne[0] : 0;
+    const int  ne1 = dst ? dst->ne[1] : 0;
+    const int  ne2 = dst ? dst->ne[2] : 0;
+    const int  ne3 = dst ? dst->ne[3] : 0;
+
+    const cl_ulong  nb0 = dst ? dst->nb[0] : 0;
+    const cl_ulong  nb1 = dst ? dst->nb[1] : 0;
+    const cl_ulong  nb2 = dst ? dst->nb[2] : 0;
+    const cl_ulong  nb3 = dst ? dst->nb[3] : 0;
+
+    GGML_ASSERT(ne10 % ne02 == 0);
+    GGML_ASSERT(ne10 >= ne02);
+
+    int nth = MIN(64, ne00);
+
+    const int n_past     = ((int *) dst->op_params)[0];
+    const int n_dims     = ((int *) dst->op_params)[1];
+    const int mode       = ((int *) dst->op_params)[2];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+    int32_t sections[4];
+
+    memcpy(&freq_base,   (int32_t *) dst->op_params + 5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params + 6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params + 7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params + 9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int32_t)*4);
+
+    const bool is_neox = mode & 2;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    const int  is_imrope = mode == GGML_ROPE_TYPE_IMROPE;
+
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
+
+    cl_kernel kernel;
+
+    if (is_neox) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_neox_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_neox_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    } else if (is_mrope && !is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_multi_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_multi_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    } else if (is_vision) {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_vision_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_vision_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        }
+    } else {
+        switch (src0->type) {
+            case GGML_TYPE_F32:
+                kernel = backend_ctx->kernel_rope_norm_f32;
+                break;
+            case GGML_TYPE_F16:
+                kernel = backend_ctx->kernel_rope_norm_f16;
+                break;
+            default:
+                GGML_ASSERT(false);
+        };
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   extra2 ? &extra2->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(int),      &ne3));
+    CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &nb0));
+    CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &nb3));
+    CL_CHECK(clSetKernelArg(kernel, 24, sizeof(int),      &n_past));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(int),      &n_dims));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &n_ctx_orig));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &freq_base));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(float),    &freq_scale));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &ext_factor));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(float),    &attn_factor));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(float),    &beta_fast));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(float),    &beta_slow));
+    // both mrope and vision kernels have sections
+    if (is_mrope || is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 33, sizeof(int32_t)*4, &sections));
+    }
+    // only mrope has is_imrope
+    if (is_mrope && !is_vision) {
+        CL_CHECK(clSetKernelArg(kernel, 34, sizeof(int), &is_imrope));
+    }
+
+    size_t global_work_size[] = {(size_t)ne01*nth, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_solve_tri(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel = backend_ctx->kernel_solve_tri_f32;
+    GGML_ASSERT(kernel != nullptr);
+
+    const int n = src0->ne[0];
+    const int k = src1->ne[0];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb10 = src1->nb[0];
+    const cl_ulong nb11 = src1->nb[1];
+    const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
+
+    const cl_ulong nb0 = dst->nb[0];
+    const cl_ulong nb1 = dst->nb[1];
+    const cl_ulong nb2 = dst->nb[2];
+    const cl_ulong nb3 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &n));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &k));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),&nb13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb3));
+
+    size_t global_work_size[3]= { (size_t)k, (size_t)dst->ne[2], (size_t)dst->ne[3]};
+    size_t local_work_size[] = {16, 4, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_im2col(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    // src0 - filter, src1 - input
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const cl_long IC = src1->ne[is_2D ? 2 : 1];
+    const cl_long IH = is_2D ? src1->ne[1] : 1;
+    const cl_long IW =         src1->ne[0];
+
+    const cl_long KH = is_2D ? src0->ne[1] : 1;
+    const cl_long KW =         src0->ne[0];
+
+    const cl_long OH = is_2D ? dst->ne[2] : 1;
+    const cl_long OW =         dst->ne[1];
+
+    // nb is byte offset, src is type float32
+    const cl_ulong delta_offset = src1->nb[is_2D ? 2 : 1]/4;
+    const cl_long  batch        = src1->ne[is_2D ? 3 : 2];
+    const cl_ulong batch_offset = src1->nb[is_2D ? 3 : 2]/4;
+
+    const cl_long pelements = OW*KW*KH;
+    const cl_long CHW       = IC*KH*KW;
+
+    cl_kernel kernel;
+
+    if(dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_im2col_f16;
+    } else {
+        kernel = backend_ctx->kernel_im2col_f32;
+    }
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),   &extra1->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(cl_ulong), &batch_offset));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(cl_ulong), &delta_offset));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(cl_long),  &IW));
+    CL_CHECK(clSetKernelArg(kernel,   7, sizeof(cl_long),  &IH));
+    CL_CHECK(clSetKernelArg(kernel,   8, sizeof(cl_long),  &IC));
+    CL_CHECK(clSetKernelArg(kernel,   9, sizeof(cl_long),  &OW));
+    CL_CHECK(clSetKernelArg(kernel,  10, sizeof(cl_long),  &OH));
+    CL_CHECK(clSetKernelArg(kernel,  11, sizeof(cl_long),  &KW));
+    CL_CHECK(clSetKernelArg(kernel,  12, sizeof(cl_long),  &KH));
+    CL_CHECK(clSetKernelArg(kernel,  13, sizeof(cl_long),  &pelements));
+    CL_CHECK(clSetKernelArg(kernel,  14, sizeof(cl_long),  &CHW));
+    CL_CHECK(clSetKernelArg(kernel,  15, sizeof(int),      &s0));
+    CL_CHECK(clSetKernelArg(kernel,  16, sizeof(int),      &s1));
+    CL_CHECK(clSetKernelArg(kernel,  17, sizeof(int),      &p0));
+    CL_CHECK(clSetKernelArg(kernel,  18, sizeof(int),      &p1));
+    CL_CHECK(clSetKernelArg(kernel,  19, sizeof(int),      &d0));
+    CL_CHECK(clSetKernelArg(kernel,  20, sizeof(int),      &d1));
+
+    const int num_blocks = (pelements + 256 - 1) / 256;
+    size_t global_work_size[] = {(size_t)num_blocks*256, (size_t)OH, (size_t)batch*IC};
+    size_t local_work_size[] = {256, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_argsort(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00  = src0->ne[0];
+    const int nrows = ggml_nrows(src0);
+
+    int ne00_padded = 1;
+    while (ne00_padded < ne00) {
+        ne00_padded *= 2;
+    }
+
+    int order = (enum ggml_sort_order) dst->op_params[0];
+
+    cl_kernel kernel = backend_ctx->kernel_argsort_f32_i32;
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),            &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong),          &offset0));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),            &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong),          &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(int),               &ne00));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(int),               &ne00_padded));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(int),               &order));
+    CL_CHECK(clSetKernelArg(kernel,   7, ne00_padded*sizeof(int),   NULL));
+
+    size_t global_work_size[] = {(size_t)ne00_padded, (size_t)nrows, (size_t)1};
+    size_t local_work_size[] = {(size_t)ne00_padded, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_sum_rows(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_UNUSED(src1);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const cl_ulong nb1  = dst->nb[1];
+    const cl_ulong nb2  = dst->nb[2];
+    const cl_ulong nb3  = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_sum_rows_f32;
+
+    CL_CHECK(clSetKernelArg(kernel,   0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,   2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,   3, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,   4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel,   5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel,   6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel,   7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel,   8, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,   9, sizeof(cl_ulong), &nb02));
+    CL_CHECK(clSetKernelArg(kernel,  10, sizeof(cl_ulong), &nb03));
+    CL_CHECK(clSetKernelArg(kernel,  11, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel,  12, sizeof(cl_ulong), &nb2));
+    CL_CHECK(clSetKernelArg(kernel,  13, sizeof(cl_ulong), &nb3));
+
+    size_t global_work_size[] = {(size_t)ne01, (size_t)ne02, (size_t)ne03};
+    size_t local_work_size[] = {(size_t)64, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+static void ggml_cl_glu(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    GGML_ASSERT(ggml_is_contiguous_1(src0));
+
+    if (src1) {
+        GGML_ASSERT(src1);
+        GGML_ASSERT(src1->extra);
+        GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    cl_kernel kernel;
+    switch (ggml_get_glu_op(dst)) {
+        case GGML_GLU_OP_GEGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu;
+            } else {
+                kernel = backend_ctx->kernel_geglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_REGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_reglu;
+            } else {
+                kernel = backend_ctx->kernel_reglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_SWIGLU:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_swiglu;
+            } else {
+                kernel = backend_ctx->kernel_swiglu_f16;
+            }
+            break;
+        case GGML_GLU_OP_SWIGLU_OAI:
+            kernel = backend_ctx->kernel_swiglu_oai;
+            break;
+        case GGML_GLU_OP_GEGLU_ERF:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_erf;
+            } else {
+                kernel = backend_ctx->kernel_geglu_erf_f16;
+            }
+            break;
+        case GGML_GLU_OP_GEGLU_QUICK:
+            if (dst->type == GGML_TYPE_F32) {
+                kernel = backend_ctx->kernel_geglu_quick;
+            } else {
+                kernel = backend_ctx->kernel_geglu_quick_f16;
+            }
+            break;
+        default:
+            GGML_ABORT("Unsupported glu op");
+    }
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    ggml_tensor_extra_cl * extra1 = src1 ? (ggml_tensor_extra_cl *)src1->extra : nullptr;
+
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    cl_ulong offset1 = extra1 ? extra1->offset + src1->view_offs : offset0;
+
+    const int ne0       = dst->ne[0];
+
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb11 = src1 ? src1->nb[1] : nb01;
+
+    const cl_ulong nb1  = dst->nb[1];
+
+    const int   swp   = ggml_get_op_params_i32(dst, 1);
+    const float alpha = ggml_get_op_params_f32(dst, 2);
+    const float limit = ggml_get_op_params_f32(dst, 3);
+
+    const int ne00_off = src1 ? 0 : (swp ? ne0 : 0);
+    const int ne10_off = src1 ? 0 : (swp ? 0 : ne0);
+
+    CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+    CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   src1 ? &extra1->data_device : &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+    CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb11));
+    CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne0));
+    CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne00_off));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne10_off));
+
+    if (ggml_get_glu_op(dst) == GGML_GLU_OP_SWIGLU_OAI) {
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float), &limit));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float), &alpha));
+    }
+
+    const size_t nrows = ggml_nrows(src0);
+    size_t nth = 512;
+    size_t global_work_size[] = {nrows*nth, 1, 1};
+    size_t local_work_size[] = {nth, 1, 1};
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+}
+
+//------------------------------------------------------------------------------
+// Op offloading
+//------------------------------------------------------------------------------
+
+typedef void (*ggml_cl_func_t)(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+
+bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor) {
+    ggml_cl_func_t func = nullptr;
+
+    ggml_tensor * src0 = tensor->src[0];
+    ggml_tensor * src1 = tensor->src[1];
+
+    const bool any_on_device = tensor->extra
+        || (src0 != nullptr && src0->extra)
+        || (src1 != nullptr && src1->extra);
+
+    switch (tensor->op) {
+        case GGML_OP_GET_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_get_rows;
+            break;
+        case GGML_OP_SET_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_set_rows;
+            break;
+        case GGML_OP_CPY:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_cpy;
+            break;
+        case GGML_OP_DUP:
+        case GGML_OP_CONT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_dup;
+            break;
+        case GGML_OP_ADD:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_add;
+            break;
+        case GGML_OP_ADD_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_add_id;
+            break;
+        case GGML_OP_MUL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul;
+            break;
+        case GGML_OP_DIV:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_div;
+            break;
+        case GGML_OP_SUB:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sub;
+            break;
+        case GGML_OP_SQR:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqr;
+            break;
+        case GGML_OP_SQRT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sqrt;
+            break;
+        case GGML_OP_MEAN:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mean;
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_GELU:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu;
+                    break;
+                case GGML_UNARY_OP_GELU_ERF:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_erf;
+                    break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_gelu_quick;
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_silu;
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_relu;
+                    break;
+                case GGML_UNARY_OP_SIGMOID:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_sigmoid;
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_tanh;
+                    break;
+                case GGML_UNARY_OP_EXPM1:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_expm1;
+                    break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_softplus;
+                    break;
+                default:
+                    return false;
+            } break;
+        case GGML_OP_GLU:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_glu;
+            break;
+        case GGML_OP_TRI:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_tri;
+            break;
+        case GGML_OP_FILL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_fill;
+            break;
+        case GGML_OP_CLAMP:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_clamp;
+            break;
+        case GGML_OP_NORM:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_norm;
+            break;
+        case GGML_OP_RMS_NORM:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_rms_norm;
+            break;
+        case GGML_OP_GROUP_NORM:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_group_norm;
+            break;
+                case GGML_OP_REPEAT:
+             if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_repeat;
+            break;
+        case GGML_OP_PAD:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_pad(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_UPSCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_upscale(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_CONV_2D:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_conv_2d;
+            break;
+        case GGML_OP_SSM_CONV:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_ssm_conv;
+            break;
+        case GGML_OP_CONCAT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_concat;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_timestep_embedding(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_MUL_MAT:
+            if (!any_on_device && !ggml_cl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
+                return false;
+            }
+            func = ggml_cl_mul_mat;
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_mul_mat_id;
+            break;
+        case GGML_OP_SCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_scale;
+            break;
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_nop;
+            break;
+        case GGML_OP_DIAG_MASK_INF:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_diag_mask_inf;
+            break;
+        case GGML_OP_SOFT_MAX:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_soft_max;
+            break;
+        case GGML_OP_ROPE:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_rope;
+            break;
+        case GGML_OP_SOLVE_TRI:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_solve_tri;
+            break;
+        case GGML_OP_IM2COL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_im2col;
+            break;
+        case GGML_OP_ARGSORT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_argsort;
+            break;
+        case GGML_OP_SUM_ROWS:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_sum_rows;
+            break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_flash_attn(backend, tensor->src[0], tensor->src[1], tensor);
+            return true;
+        default:
+            return false;
+    }
+
+    func(backend, tensor->src[0], tensor->src[1], tensor);
+    return true;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl
new file mode 100644
index 0000000..509bf17
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/add.cl
@@ -0,0 +1,190 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add
+//------------------------------------------------------------------------------
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+kernel void kernel_add(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) + *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        global float4 * src0,
+        ulong  offset0,
+        global float4 * src1,
+        ulong  offset1,
+        global float4 * dst,
+        ulong  offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] + src1[idx1];
+}
+
+kernel void kernel_add_f16(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global char * dst,
+        ulong  offsetd,
+        int   ne00,
+        int   ne01,
+        int   ne02,
+        int   ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int   ne10,
+        int   ne11,
+        int   ne12,
+        int   ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int   ne0,
+        int   ne1,
+        int   ne2,
+        int   ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int type_src0,
+        int type_src1
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+
+        half v0, v1;
+        if (type_src0 == 1) {
+            v0 = convert_half(*((global float *)(src0_ptr + i0*nb00)));
+        } else {
+            v0 = *((global half *)(src0_ptr + i0*nb00));
+        }
+
+        if (type_src1 == 1) {
+            v1 = convert_half(*((global float *)(src1_ptr + i10*nb10)));
+        } else {
+            v1 = *((global half *)(src1_ptr + i10*nb10));
+        }
+
+        *((global half *)(dst_ptr + i0*nb0)) = v0 + v1;
+    }
+}
+
+kernel void kernel_add_row_f16(
+        global char * src0,
+        ulong  offset0,
+        global char * src1,
+        ulong  offset1,
+        global half4 * dst,
+        ulong  offsetd,
+        int ne,
+        int type_src0,
+        int type_src1
+) {
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+
+    half4 v0, v1;
+    if (type_src0 == 1) {
+        global float4* src0_f32 = (global float4*)((global char*)src0 + offset0);
+        v0 = convert_half4(src0_f32[gid]);
+    } else {
+        global half4* src0_f16 = (global half4*)((global char*)src0 + offset0);
+        v0 = src0_f16[gid];
+    }
+
+    if (type_src1 == 1) {
+        global float4* src1_f32 = (global float4*)((global char*)src1 + offset1);
+        v1 = convert_half4(src1_f32[idx1]);
+    } else {
+        global half4* src1_f16 = (global half4*)((global char*)src1 + offset1);
+        v1 = src1_f16[idx1];
+    }
+
+    dst[gid] = v0 + v1;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl
new file mode 100644
index 0000000..e9c6d55
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/add_id.cl
@@ -0,0 +1,42 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// add_id
+//------------------------------------------------------------------------------
+kernel void kernel_add_id(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb11,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    src2 = (global char*)((global char*)src2 + offset2);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int i1 = get_group_id(0);
+    int i2 = get_group_id(1);
+
+    const int i11 = *((global const int *) (src2 + i1*sizeof(int) + i2*nb21));
+
+    const size_t nb1 = ne0 * sizeof(float);
+    const size_t nb2 = ne1 * nb1;
+
+    global float * dst_row  = (global float *)((global char *)dst  + i1*nb1 + i2*nb2);
+    global float * src0_row = (global float *)((global char *)src0 + i1*nb01 + i2*nb02);
+    global float * src1_row = (global float *)((global char *)src1 + i11*nb11);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        dst_row[i0] = src0_row[i0] + src1_row[i0];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl
new file mode 100644
index 0000000..af4adc7
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/argsort.cl
@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define SWAP(x, y, T) { T tmp = (x); (x) = (y); (y) = tmp; }
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+kernel void kernel_argsort_f32_i32(
+    global float * src0,
+    ulong          offset0,
+    global int   * dst,
+    ulong          offsetd,
+    const int      ne00,
+    const int      ne00_pad,
+    const int      order,
+    local int    * dst_row
+) {
+    // bitonic sort
+    int col = get_local_id(0);
+    int row = get_group_id(1);
+
+    if (col >= ne00_pad) {
+        return;
+    }
+
+    src0 = (global char  *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    global float * x_row = src0 + row * ne00;
+
+    // initialize indices
+    dst_row[col] = col;
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    for (int k = 2; k <= ne00_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ne00 ||
+                        (dst_row[ixj] < ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj], int);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ne00 ||
+                        (dst_row[col] < ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj], int);
+                    }
+                }
+            }
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ne00) {
+        dst[row * ne00 + col] = dst_row[col];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl
new file mode 100644
index 0000000..ae60324
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/clamp.cl
@@ -0,0 +1,20 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// clamp
+//------------------------------------------------------------------------------
+kernel void kernel_clamp(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float min,
+        float max
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = src0[get_global_id(0)] < min ?
+        min :
+        (src0[get_global_id(0)] > max ? max : src0[get_global_id(0)]);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl
new file mode 100644
index 0000000..0c1b3d7
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/concat.cl
@@ -0,0 +1,51 @@
+kernel void kernel_concat_f32(
+    global  const char * src0,
+    ulong                offset0,
+    global  const char * src1,
+    ulong                offset1,
+    global        char * dst,
+    ulong                offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb00,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb10,
+    ulong           nb11,
+    ulong           nb12,
+    ulong           nb13,
+    int             ne0,
+    ulong           nb0,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3,
+    int             dim
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    int o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+    global const float * x;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (global const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (global const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+        }
+
+        global float * y = (global float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl
new file mode 100644
index 0000000..e339c90
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d.cl
@@ -0,0 +1,185 @@
+#ifdef USE_FP16
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#define T_FLOAT half
+#define T_FLOAT4 half4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore_half4_rte(data, offset, p)
+#else
+#define T_FLOAT float
+#define T_FLOAT4 float4
+#define VSTORE_T_FLOAT4(data, offset, p) vstore4(data, offset, p)
+#endif
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global T_FLOAT* knl_data = (global T_FLOAT*) ((global char*)p_knl + off_knl);
+    global T_FLOAT* src_data = (global T_FLOAT*) ((global char*)p_src + off_src);
+    global T_FLOAT* dst_data = (global T_FLOAT*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local T_FLOAT* Ash = (local T_FLOAT*)shared;
+    local T_FLOAT4* Bsh = (local T_FLOAT4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (T_FLOAT)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            T_FLOAT4 val = (T_FLOAT4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((T_FLOAT*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            T_FLOAT regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                T_FLOAT4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), convert_float4(regB), regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                VSTORE_T_FLOAT4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = (T_FLOAT)(((float*)&res)[v]);
+                    }
+                }
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl
new file mode 100644
index 0000000..cb05637
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/conv2d_f16_f32.cl
@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define T_ACCUM float4
+#define VEC_SIZE 4
+
+#define BS_K 64
+#define BS_NPQ 64
+#define BS_CRS 16
+
+#define TS_K 4
+#define TS_NPQ 8
+
+#define WG_K (BS_K / TS_K)
+#define WG_NPQ (BS_NPQ / TS_NPQ)
+
+#define BS_NPQ_VEC (BS_NPQ / VEC_SIZE)
+#define TS_NPQ_VEC (TS_NPQ / VEC_SIZE)
+
+static inline uint splitWork(uint work_size, uint block_size){
+    return (work_size + block_size - 1) / block_size;
+}
+
+REQD_SUBGROUP_SIZE_128
+kernel void kernel_conv_2d(
+    global void* p_knl,
+    ulong off_knl,
+    global void* p_src,
+    ulong off_src,
+    global void* p_dst,
+    ulong off_dst,
+    local void* shared,
+    uint Cout, uint Cin, uint N,
+    uint KW, uint KH, uint W, uint H, uint OW, uint OH,
+    uint s0, uint s1, uint p0, uint p1, uint d0, uint d1,
+    uint nb01, uint nb02, uint nb03,
+    uint nb11, uint nb12, uint nb13,
+    uint nb1, uint nb2, uint nb3
+) {
+    global half* knl_data = (global half*) ((global char*)p_knl + off_knl);
+    global float* src_data = (global float*) ((global char*)p_src + off_src);
+    global float* dst_data = (global float*) ((global char*)p_dst + off_dst);
+
+    const uint K = Cout;
+    const uint CRS = Cin*KH*KW;
+    const uint NPQ = N*OH*OW;
+
+    const uint lid_k = get_local_id(0);
+    const uint lid_npq = get_local_id(1);
+    const uint tid = lid_npq * WG_K + lid_k;
+
+    const uint B_idx_K = get_group_id(0);
+    const uint B_idx_NPQ = get_group_id(1);
+
+    const uint offset_k = B_idx_K * BS_K;
+    const uint offset_npq = B_idx_NPQ * BS_NPQ;
+
+    local half* Ash = (local half*)shared;
+    local float4* Bsh = (local float4*) &Ash[BS_K * BS_CRS];
+
+    T_ACCUM regC[TS_K][TS_NPQ_VEC];
+    for (int i = 0; i < TS_K; ++i) {
+        for (int j = 0; j < TS_NPQ_VEC; ++j) {
+            regC[i][j] = (T_ACCUM)(0.0f);
+        }
+    }
+
+    const uint NB_CRS = splitWork(CRS, BS_CRS);
+
+    for (uint B_idx_CRS = 0; B_idx_CRS < NB_CRS; ++B_idx_CRS) {
+        const uint offset_crs = B_idx_CRS * BS_CRS;
+
+        for (int i = tid; i < BS_K * BS_CRS; i += (WG_K * WG_NPQ)) {
+            const uint k_l = i / BS_CRS;
+            const uint crs_l = i % BS_CRS;
+            const uint k_g = offset_k + k_l;
+            const uint crs_g = offset_crs + crs_l;
+
+            if (k_g < K && crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW*KH);
+                const uint KH_idx = (crs_g - Cin_idx*KW*KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx*KW*KH - KH_idx*KW;
+                const uint knl_idx = KW_idx + KH_idx*nb01 + Cin_idx*nb02 + k_g*nb03;
+                Ash[k_l * BS_CRS + crs_l] = knl_data[knl_idx];
+            } else {
+                Ash[k_l * BS_CRS + crs_l] = (half)0.0f;
+            }
+        }
+
+        for (int i = tid; i < BS_CRS * BS_NPQ_VEC; i += (WG_K * WG_NPQ)) {
+            const uint crs_l = i / BS_NPQ_VEC;
+            const uint npq_l_vec = i % BS_NPQ_VEC;
+            const uint crs_g = offset_crs + crs_l;
+
+            float4 val = (float4)(0.0f);
+            if (crs_g < CRS) {
+                const uint Cin_idx = crs_g / (KW * KH);
+                const uint KH_idx = (crs_g - Cin_idx * KW * KH) / KW;
+                const uint KW_idx = crs_g - Cin_idx * KW * KH - KH_idx * KW;
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = offset_npq + npq_l_vec * VEC_SIZE + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx = npq_g / (OH * OW);
+                        const uint pq_idx = npq_g % (OH * OW);
+                        const uint OH_idx = pq_idx / OW;
+                        const uint OW_idx = pq_idx % OW;
+                        const int H_idx = (int)(OH_idx * s1 + KH_idx * d1 - p1);
+                        const int W_idx = (int)(OW_idx * s0 + KW_idx * d0 - p0);
+
+                        if (H_idx >= 0 && H_idx < H && W_idx >= 0 && W_idx < W) {
+                            const uint src_idx = W_idx + H_idx * nb11 + Cin_idx * nb12 + N_idx * nb13;
+                            ((float*)&val)[v] = src_data[src_idx];
+                        }
+                    }
+                }
+            }
+            Bsh[crs_l * BS_NPQ_VEC + npq_l_vec] = val;
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (uint crs_l = 0; crs_l < BS_CRS; ++crs_l) {
+            half regA[TS_K];
+            for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                regA[k_l_reg] = Ash[(lid_k * TS_K + k_l_reg) * BS_CRS + crs_l];
+            }
+
+            for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+                float4 regB = Bsh[crs_l * BS_NPQ_VEC + lid_npq * TS_NPQ_VEC + npq_l_vec_reg];
+                for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+                    regC[k_l_reg][npq_l_vec_reg] = mad(convert_float(regA[k_l_reg]), regB, regC[k_l_reg][npq_l_vec_reg]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (uint k_l_reg = 0; k_l_reg < TS_K; ++k_l_reg) {
+        const uint k_g = offset_k + lid_k * TS_K + k_l_reg;
+        if (k_g >= K) continue;
+
+        for (uint npq_l_vec_reg = 0; npq_l_vec_reg < TS_NPQ_VEC; ++npq_l_vec_reg) {
+            const uint npq_g_base = offset_npq + (lid_npq * TS_NPQ_VEC + npq_l_vec_reg) * VEC_SIZE;
+
+            const uint N_idx = npq_g_base / (OH * OW);
+            const uint pq_idx = npq_g_base % (OH * OW);
+            const uint OH_idx = pq_idx / OW;
+            const uint OW_idx = pq_idx % OW;
+
+            if (nb1 == OW && OW_idx + VEC_SIZE <= OW && npq_g_base + VEC_SIZE <= NPQ) {
+                const uint dst_idx = OW_idx + OH_idx*nb1 + k_g*nb2 + N_idx*nb3;
+                vstore4(regC[k_l_reg][npq_l_vec_reg], 0, &dst_data[dst_idx]);
+            } else {
+                T_ACCUM res = regC[k_l_reg][npq_l_vec_reg];
+                for (int v = 0; v < VEC_SIZE; ++v) {
+                    const uint npq_g = npq_g_base + v;
+                    if (npq_g < NPQ) {
+                        const uint N_idx_s = npq_g / (OH*OW);
+                        const uint pq_idx_s = npq_g % (OH*OW);
+                        const uint OH_idx_s = pq_idx_s / OW;
+                        const uint OW_idx_s = pq_idx_s % OW;
+                        const uint dst_idx_s = OW_idx_s + OH_idx_s*nb1 + k_g*nb2 + N_idx_s*nb3;
+                        dst_data[dst_idx_s] = ((float*)&res)[v];
+                    }
+                }
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl
new file mode 100644
index 0000000..9369351
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/cpy.cl
@@ -0,0 +1,184 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// cpy
+//------------------------------------------------------------------------------
+
+kernel void kernel_cpy_f16_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f16_f32(
+        global half * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global half * src = (global half *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f16(
+        global float * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global half * dst_data = (global half *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
+
+kernel void kernel_cpy_f32_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    int i3 = n / (ne2*ne1*ne0);
+    int i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    int i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    int i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    global float * dst_data = (global float *) ((global char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        global const float * src = (global float *)((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        dst_data[i00] = src[0];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl
new file mode 100644
index 0000000..9fb4347
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/cvt.cl
@@ -0,0 +1,366 @@
+//------------------------------------------------------------------------------
+// This file is contains kernels for data conversion.
+// These kernels are used when loading the model, so its performance is less
+// important.
+//------------------------------------------------------------------------------
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+struct block_q6_K {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;                  // super-block scale
+};
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_0
+// Convert the block_q4_0 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_0(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK4_0/2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q4_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_0_noshuffle
+// Flatten q4_0 weights and unshuffle the bits
+//------------------------------------------------------------------------------
+
+kernel void kernel_convert_block_q4_0_noshuffle(
+    global struct block_q4_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+    for (int i = 0; i < QK4_0/4; ++i) {
+        uchar x0 = b->qs[2*i + 0];
+        uchar x1 = b->qs[2*i + 1];
+
+        q[i + 0      ] = convert_uchar(x0 & 0x0F) | convert_uchar((x1 & 0x0F) << 4);
+        q[i + QK4_0/4] = convert_uchar((x0 & 0xF0) >> 4) | convert_uchar(x1 & 0xF0);
+
+#ifdef ADRENO_GPU
+        // Workaround for adreno - must have the following printf statement for
+        // the kernel to work properly. Otherwise it produces incorrect result.
+        // convert_uchar above also seems necessary.
+        // Compare against a large number so that it does not print anything.
+        // get_sub_group_local_id() also works.
+        if (get_global_id(0) == 65536*4096) {
+            printf("%04x - %02x\n", *(global ushort*)d, ((x0 & 0xF0) >> 4) | (x1 & 0xF0));
+        }
+#endif
+    }
+}
+
+kernel void kernel_restore_block_q4_0_noshuffle(
+    global uchar * src_q,
+    global half  * src_d,
+    global struct block_q4_0 * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q4_0 * b = (global struct block_q4_0 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK4_0/2*get_global_id(0);
+    global half  * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK4_0/4; ++i) {
+        uchar x0 = q[i + 0      ] ;
+        uchar x1 = q[i + QK4_0/4];
+
+        b->qs[2*i + 0] = convert_uchar((x0 & mask_0F) | ((x1 & mask_0F) << 4));
+        b->qs[2*i + 1] = convert_uchar(((x0 & mask_F0) >> 4) | (x1 & mask_F0));
+    }
+}
+
+//------------------------------------------------------------------------------
+// block_mxfp4
+//------------------------------------------------------------------------------
+#define QK_MXFP4 32
+struct block_mxfp4 {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4 / 2];
+};
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_mxfp4
+// Convert the block_mxfp4 format to 2 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_mxfp4(
+    global struct block_mxfp4 * src0,
+    global uchar * dst_q,
+    global uchar * dst_e
+) {
+    global struct block_mxfp4 * b = (global struct block_mxfp4 *) src0 + get_global_id(0);
+    global uchar * q = (global uchar *) dst_q + QK_MXFP4 / 2 * get_global_id(0);
+    global uchar * e = (global uchar *) dst_e + get_global_id(0);
+
+    *e = b->e;
+
+    for (int i = 0; i < QK_MXFP4 / 2; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_convert_block_mxfp4_trans(
+    global struct block_mxfp4 * src0,
+    __global uint4 * dst_q,
+    __global uchar * dst_e,
+    uint ne00,
+    uint ne01
+) {
+    int i00 = get_global_id(1);
+    uint i01 = get_global_id(0);
+    uint i02 = get_global_id(2);
+
+    uint ne00_blk = ne00 / QK_MXFP4;
+    uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+    uint dst_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
+
+    global struct block_mxfp4 * b = src0 + src_blk_offset;
+
+    dst_q[dst_blk_offset] = ((global uint4 *)(&(b->qs[0])))[0];
+    dst_e[dst_blk_offset] = b->e;
+}
+
+kernel void kernel_restore_block_mxfp4(
+    global uchar * src_q,
+    global half  * src_e,
+    global struct block_mxfp4 * dst
+) {
+    global struct block_mxfp4 * b = (global struct block_mxfp4 *) dst + get_global_id(0);
+    global uchar * q = (global uchar *) src_q + QK_MXFP4 / 2 * get_global_id(0);
+    global uchar * e = (global uchar *) src_e + get_global_id(0);
+
+    b->e = *e;
+    for (int i = 0; i < QK_MXFP4 / 2; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+kernel void kernel_restore_block_mxfp4_trans(
+    __global uint4 * src_q,
+    __global uchar * src_e,
+    global struct block_mxfp4 * dst,
+    uint ne00,
+    uint ne01
+) {
+    int i00 = get_global_id(1);
+    uint i01 = get_global_id(0);
+    uint i02 = get_global_id(2);
+
+    uint ne00_blk = ne00 / QK_MXFP4;
+    uint src_blk_offset = i01 + i00 * ne01 + i02 * ne00_blk * ne01;
+    uint dst_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+
+    global struct block_mxfp4 * b = dst + dst_blk_offset;
+
+    ((global uint4 *)(&(b->qs[0])))[0] = src_q[src_blk_offset];
+    b->e = src_e[src_blk_offset];
+}
+
+//------------------------------------------------------------------------------
+// block_q8_0
+//------------------------------------------------------------------------------
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+kernel void kernel_convert_block_q8_0(
+    global block_q8_0 * src0,
+    global uchar * dst_q,
+    global half  * dst_d
+) {
+    global block_q8_0 * b = (global block_q8_0 *) src0 + get_global_id(0);
+    global uchar      * q = (global uchar *) dst_q + QK8_0*get_global_id(0);
+    global half       * d = (global half *) dst_d + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK8_0; ++i) {
+        q[i] = b->qs[i];
+    }
+}
+
+kernel void kernel_restore_block_q8_0(
+    global uchar * src_q,
+    global half  * src_d,
+    global block_q8_0 * dst
+) {
+    global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0);
+    global uchar      * q = (global uchar *) src_q + QK8_0*get_global_id(0);
+    global half       * d = (global half *) src_d + get_global_id(0);
+
+    b->d = *d;
+    for (int i = 0; i < QK8_0; ++i) {
+        b->qs[i] = q[i];
+    }
+}
+
+kernel void kernel_restore_block_q8_0_trans(
+    global uchar * src_q,
+    global half  * src_d,
+    global block_q8_0 * dst,
+    uint ne00,
+    uint ne01
+){
+    uint num_blk_per_row = ne00 / QK8_0;
+
+    global block_q8_0 * b = (global block_q8_0 *) dst + get_global_id(0) * num_blk_per_row;
+    global uchar      * q = (global uchar *) src_q + get_global_id(0) * 4; // 4 8-bit packed
+    global half       * d = (global half *) src_d + get_global_id(0);
+
+    for (uint blk = 0; blk < num_blk_per_row; blk++) {
+        b->d = *d;
+
+        for (uint i = 0; i < QK8_0; i+=4) {
+            b->qs[i]   = q[0];
+            b->qs[i+1] = q[1];
+            b->qs[i+2] = q[2];
+            b->qs[i+3] = q[3];
+
+            q += 4 * ne01; // M stride
+        }
+
+        d += ne01;
+
+        b++;
+    }
+}
+
+//------------------------------------------------------------------------------
+// kernel_convert_block_q6_K
+// Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).
+// This kernel does not deshuffle the bits.
+// Each thread processes a super block.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q6_K(
+    global struct block_q6_K * src0,
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) src0 + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    *d = b->d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        ql[i] = b->ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        qh[i] = b->qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        s[i] = b->scales[i];
+    }
+}
+
+// Restore block_q6_K from flattened arrays.
+// Each thread processes a super block.
+kernel void kernel_restore_block_q6_K(
+    global uchar * dst_ql,
+    global uchar * dst_qh,
+    global char  * dst_s,
+    global half  * dst_d,
+    global struct block_q6_K * dst
+) {
+    global struct block_q6_K * b = (global struct block_q6_K *) dst + get_global_id(0);
+    global uchar * ql = (global uchar *) dst_ql + QK_K/2*get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/4*get_global_id(0);
+    global char  * s  = (global char  *) dst_s  + QK_K/16*get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+
+    b->d = *d;
+
+    for (int i = 0; i < QK_K/2; ++i) {
+        b->ql[i] = ql[i];
+    }
+    for (int i = 0; i < QK_K/4; ++i) {
+        b->qh[i] = qh[i];
+    }
+    for (int i = 0; i < QK_K/16; ++i) {
+        b->scales[i] = s[i];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl
new file mode 100644
index 0000000..36eff04
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/diag_mask_inf.cl
@@ -0,0 +1,58 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// diag_mask_inf kernels
+//------------------------------------------------------------------------------
+kernel void kernel_diag_mask_inf(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i02 = get_global_id(2);
+    int i01 = get_global_id(1);
+    int i00 = get_global_id(0);
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_past
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    int i = 2*get_global_id(0);
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int i4 = 4*i;
+    int i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    int i01 = i4/(ne00);      i4 -= i01*ne00;
+    int i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        (&dst[i+1])[k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            (&dst[i])[k] = -INFINITY;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl
new file mode 100644
index 0000000..6d9b4ad
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/div.cl
@@ -0,0 +1,138 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// div
+//------------------------------------------------------------------------------
+kernel void kernel_div(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) / *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_div_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] / src1[idx1];
+}
+
+kernel void kernel_div_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) / *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_div_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] / src1[idx1];
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py b/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py
new file mode 100644
index 0000000..b5d1d72
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/embed_kernel.py
@@ -0,0 +1,26 @@
+#
+
+import sys
+import logging
+logger = logging.getLogger("opencl-embed-kernel")
+
+
+def main():
+    logging.basicConfig(level=logging.INFO)
+
+    if len(sys.argv) != 3:
+        logger.info("Usage: python embed_kernel.py <input_file> <output_file>")
+        sys.exit(1)
+
+    ifile = open(sys.argv[1], "r")
+    ofile = open(sys.argv[2], "w")
+
+    for i in ifile:
+        ofile.write('R"({})"\n'.format(i))
+
+    ifile.close()
+    ofile.close()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl
new file mode 100644
index 0000000..126298a
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/expm1.cl
@@ -0,0 +1,82 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// expm1
+//------------------------------------------------------------------------------
+kernel void kernel_expm1_f32_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}
+
+kernel void kernel_expm1_f16_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl
new file mode 100644
index 0000000..9b73938
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/fill.cl
@@ -0,0 +1,17 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// fill
+//------------------------------------------------------------------------------
+__kernel void kernel_fill_f32(
+        __global float *dst,
+        ulong offsetd,
+        float v,
+        int n
+
+) {
+    dst = (global float*)((global char*)dst + offsetd);
+    if(get_global_id(0) < n){
+        dst[get_global_id(0)] = v;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
new file mode 100644
index 0000000..8f43c4f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f16.cl
@@ -0,0 +1,370 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE half
+#define DATA_TYPE4 half4
+#define CONVERT_ACC4(x) convert_float4(x)
+#define CONVERT_DATA4(x) convert_half4(x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
new file mode 100644
index 0000000..a6d7479
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32.cl
@@ -0,0 +1,371 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE float
+#define DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
+#define CONVERT_ACC4(x) (x)
+#define CONVERT_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
new file mode 100644
index 0000000..ec7361b
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl
@@ -0,0 +1,373 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define Q_DATA_TYPE4 float4
+#define KV_DATA_TYPE4 half4
+#define O_DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
+#define CONVERT_Q_ACC4(x) (x)
+#define CONVERT_KV_ACC4(x) convert_float4(x)
+#define CONVERT_O_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        if (sinks_void != NULL) {
+            const global ACC_TYPE* sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+            const ACC_TYPE m_sink = sinks_ptr[head_idx];
+            const ACC_TYPE m_final = max(m_i, m_sink);
+
+            const ACC_TYPE scale_o = exp(m_i - m_final);
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] *= scale_o;
+            }
+
+            l_i = l_i * exp(m_i - m_final) + exp(m_sink - m_final);
+        }
+
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3,
+    const global void* sinks_void,
+    const ulong sinks_offset
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    const global ACC_TYPE* sinks_ptr = NULL;
+    if (sinks_void != NULL) {
+        sinks_ptr = (const global ACC_TYPE*)((const global char*)sinks_void + sinks_offset);
+    }
+
+    ACC_TYPE m_i = (sinks_ptr != NULL) ? sinks_ptr[head_idx] : -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+    ACC_TYPE l_final = local_l[0];
+
+    if (sinks_ptr != NULL) {
+        l_final += exp(sinks_ptr[head_idx] - m_final);
+    }
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_O_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl
new file mode 100644
index 0000000..1ab426c
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gelu.cl
@@ -0,0 +1,89 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// gelu
+//------------------------------------------------------------------------------
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+#define SQRT_2_INV      0.70710678118654752440084436210484f
+
+kernel void kernel_gelu(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_erf(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV));
+}
+
+kernel void kernel_gelu_erf_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = 0.5f*x*(1.0f + erf(x*SQRT_2_INV));
+}
+
+kernel void kernel_gelu_quick(
+    global float * src0,
+    ulong offset0,
+    global float * dst,
+    ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * dst,
+    ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl
new file mode 100644
index 0000000..3917aa3
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemm_moe_mxfp4_f32.cl
@@ -0,0 +1,162 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#define QK_MXFP4 32
+#define N_SIMDGROUP 2
+#define SIMDGROUP_WIDTH 64
+
+static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) {
+    ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00;
+    fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00;
+    fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00;
+    fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0;
+    fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0;
+    fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0;
+    fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s0 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s0 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s0 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s0 & 0x8000;
+
+    fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0;
+    fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0;
+
+    ushort2 fp16_packed_a_1, fp16_packed_b_1;
+    fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00;
+    fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00;
+    fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00;
+    fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0;
+    fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0;
+    fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0;
+    fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s1 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s1 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s1 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s1 & 0x8000;
+
+    fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1;
+    fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1;
+
+    return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemm_moe_mxfp4_f32(
+    __global uint4 * src0_q,
+    __global uchar * src0_e,
+    __read_only image1d_buffer_t src1,
+    __global ushort4 * src2,
+    __global float * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    int           tile_size
+) {
+    uint i01  = get_global_id(0);
+    uint i20  = get_global_id(2);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+
+    ushort4 router = src2[i20];
+    ushort expert_id = router.x;
+    ushort i11 = router.y;
+    ushort i1 = router.z;
+    ushort tile_id = router.w;
+
+    if (tile_id * tile_size + i01 >= ne01) { // handle edge case when ne01 is not multiple of tile_size
+        return;
+    }
+
+    uint expert_offset = expert_id * ne00 * ne01 / 32;
+    uint tile_offset = expert_offset + tile_id * tile_size + i01;
+
+    __private float sum = 0.0f; // each thread calculate partial sum of one output
+
+    // loop along ne00 in block granularity, skip 4 blocks every iter
+    for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) {
+        // load one block of q
+        uint4 regQ = src0_q[tile_offset + ib00 * ne01];
+        // convert 8 fp4 to fp16
+        half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0));
+
+        uint offset = i11 * ne00 / 4 + ib00 * 8;
+        float4 shared_y4;
+        shared_y4 = read_imagef(src1, (offset + 0));
+        float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 4));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1));
+
+        shared_y4 = read_imagef(src1, (offset + 1));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 5));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2));
+
+        shared_y4 = read_imagef(src1, (offset + 2));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 6));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3));
+
+        shared_y4 = read_imagef(src1, (offset + 3));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 7));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+        uchar regE = src0_e[tile_offset + ib00 * ne01];
+        sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum;
+    // if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum;
+    // if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    // if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01 + tile_id * tile_size + i1 * ne01] = sum;
+    }
+
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl
new file mode 100644
index 0000000..b4b1e51
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_moe_mxfp4_f32.cl
@@ -0,0 +1,156 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#define QK_MXFP4 32
+#define N_SIMDGROUP 4
+#define SIMDGROUP_WIDTH 64
+
+static inline half8 mxfp4_to_fp16_packed8(ushort2 fp4x8) { //, ushort 0x0E00, ushort 0x8000) {
+    ushort2 fp16_packed_a_0, fp16_packed_b_0, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a_0.lo = (fp4x8.s0 << 9) & 0x0E00;
+    fp16_packed_a_0.hi = (fp4x8.s0 << 5) & 0x0E00;
+    fp16_packed_b_0.lo = (fp4x8.s0 << 1) & 0x0E00;
+    fp16_packed_b_0.hi = (fp4x8.s0 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_0.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_0.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_0.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_0.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_0.lo = (fp16_packed_a_0.lo != 0x0200) ? fp16_packed_a_0.lo : 0x0;
+    fp16_packed_a_0.hi = (fp16_packed_a_0.hi != 0x0200) ? fp16_packed_a_0.hi : 0x0;
+    fp16_packed_b_0.lo = (fp16_packed_b_0.lo != 0x0200) ? fp16_packed_b_0.lo : 0x0;
+    fp16_packed_b_0.hi = (fp16_packed_b_0.hi != 0x0200) ? fp16_packed_b_0.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s0 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s0 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s0 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s0 & 0x8000;
+
+    fp16_packed_a_0 = sign_a + bias_a + fp16_packed_a_0;
+    fp16_packed_b_0 = sign_b + bias_b + fp16_packed_b_0;
+
+    ushort2 fp16_packed_a_1, fp16_packed_b_1;
+    fp16_packed_a_1.lo = (fp4x8.s1 << 9) & 0x0E00;
+    fp16_packed_a_1.hi = (fp4x8.s1 << 5) & 0x0E00;
+    fp16_packed_b_1.lo = (fp4x8.s1 << 1) & 0x0E00;
+    fp16_packed_b_1.hi = (fp4x8.s1 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a_1.lo != 0) ? 0x3800 : 0x0;
+    bias_a.hi = (fp16_packed_a_1.hi != 0) ? 0x3800 : 0x0;
+    bias_b.lo = (fp16_packed_b_1.lo != 0) ? 0x3800 : 0x0;
+    bias_b.hi = (fp16_packed_b_1.hi != 0) ? 0x3800 : 0x0;
+
+    fp16_packed_a_1.lo = (fp16_packed_a_1.lo != 0x0200) ? fp16_packed_a_1.lo : 0x0;
+    fp16_packed_a_1.hi = (fp16_packed_a_1.hi != 0x0200) ? fp16_packed_a_1.hi : 0x0;
+    fp16_packed_b_1.lo = (fp16_packed_b_1.lo != 0x0200) ? fp16_packed_b_1.lo : 0x0;
+    fp16_packed_b_1.hi = (fp16_packed_b_1.hi != 0x0200) ? fp16_packed_b_1.hi : 0x0;
+
+    sign_a.lo = (fp4x8.s1 << 12) & 0x8000;
+    sign_a.hi = (fp4x8.s1 << 8) & 0x8000;
+    sign_b.lo = (fp4x8.s1 << 4) & 0x8000;
+    sign_b.hi = fp4x8.s1 & 0x8000;
+
+    fp16_packed_a_1 = sign_a + bias_a + fp16_packed_a_1;
+    fp16_packed_b_1 = sign_b + bias_b + fp16_packed_b_1;
+
+    return as_half8((ushort8)(fp16_packed_a_0, fp16_packed_b_0, fp16_packed_a_1, fp16_packed_b_1));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+
+__attribute__((qcom_reqd_sub_group_size("half")))
+__kernel void kernel_gemv_moe_mxfp4_f32(
+    __global uint4 * src0_q,
+    __global uchar * src0_e,
+    __read_only image1d_buffer_t src1,
+    __global uint * src2,
+    __global float * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    int           ne11
+) {
+    uint i01  = get_global_id(0);
+    uint i20  = get_global_id(2);
+    uint sgid = get_local_id(1);
+    uint slid = get_sub_group_local_id();
+
+    uint i11 = i20 % ne11;
+
+    uint expert_id = src2[i20];
+    uint expert_offset = expert_id * ne00 * ne01 / 32;
+
+    __private float sum = 0.0f; // each thread calculate partial sum of one output
+
+    // loop along ne00 in block granularity, skip 4 blocks every iter
+    for (uint ib00 = sgid; ib00 < (ne00 / QK_MXFP4); ib00 += N_SIMDGROUP) {
+
+        // load one block of q
+        uint4 regQ = src0_q[expert_offset + ib00 * ne01 + i01];
+
+        uint offset = i11 * ne00 / 4 + ib00 * 8;
+
+        half8 fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s0));
+
+        float4 shared_y4;
+        shared_y4 = read_imagef(src1, (offset + 0));
+        float4 acc = shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 4));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s1));
+
+        shared_y4 = read_imagef(src1, (offset + 1));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 5));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s2));
+
+        shared_y4 = read_imagef(src1, (offset + 2));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 6));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+
+        fp16x8 = mxfp4_to_fp16_packed8(as_ushort2(regQ.s3));
+
+        shared_y4 = read_imagef(src1, (offset + 3));
+        acc += shared_y4 * (float4)(fp16x8.s0, fp16x8.s2, fp16x8.s4, fp16x8.s6);
+
+        shared_y4 = read_imagef(src1, (offset + 7));
+        acc += shared_y4 * (float4)(fp16x8.s1, fp16x8.s3, fp16x8.s5, fp16x8.s7);
+
+        uchar regE = src0_e[ib00 * ne01 + i01 + expert_offset];
+        sum += e8m0_to_fp32(regE) * ((acc.s0 + acc.s1) + (acc.s2 + acc.s3));
+    }
+
+    // reduction in local memory, assumes #subgroups=4
+    __local float reduceLM[SIMDGROUP_WIDTH * (N_SIMDGROUP - 1)];
+    if (sgid == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = sum;
+    if (sgid == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = sum;
+    if (sgid == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (sgid == 0) sum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per thread in subgroup 0
+    if (sgid == 0) {
+        dst = dst + (offsetd >> 2);
+        dst[i01 + i20 * ne01] = sum;
+    }
+
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl
new file mode 100644
index 0000000..ee5c79f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle.cl
@@ -0,0 +1,268 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+// assume
+#define QK4_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C (0)
+        uint K,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    __private uint4     regA;
+    __private half2     regS;
+    __private float8    regB;
+
+    __private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl
new file mode 100644
index 0000000..469d3ed
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general.cl
@@ -0,0 +1,274 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+// assume
+#define QK4_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) - 8) * scale.s0 * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F) - 8) * scale.s0 * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) - 8) * scale.s0 * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) - 8) * scale.s0 * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) - 8) * scale.s0 * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) - 8) * scale.s1 * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) - 8) * scale.s1 * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) - 8) * scale.s1 * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) - 8) * scale.s1 * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) - 8) * scale.s1 * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C (0)
+        int ne00,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M / 2;
+    uint BLOCK_STRIDE_A = N_SIMDGROUP * M;
+
+    __private uint4     regA;
+    __private half2     regS;
+    __private float8    regB;
+
+    __private float2 totalSum = (float2)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    for (uint k = groupId; k < (K / QK4_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of two rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAT
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float2 reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl
new file mode 100644
index 0000000..f944ef3
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/gemv_noshuffle_general_q8_0_f32.cl
@@ -0,0 +1,195 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK8_0 32
+#define N_SIMDGROUP 4
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1(total_sums, bits8, scale, y) \
+    float shared_y; \
+    char elem; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    elem = (char)(bits8.s0 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    elem = (char)((bits8.s0 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    elem = (char)((bits8.s0 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    elem = (char)((bits8.s0 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    elem = (char)(bits8.s1 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    elem = (char)((bits8.s1 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    elem = (char)((bits8.s1 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    elem = (char)((bits8.s1 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    elem = (char)(bits8.s2 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    elem = (char)((bits8.s2 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    elem = (char)((bits8.s2 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    elem = (char)((bits8.s2 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    elem = (char)(bits8.s3 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    elem = (char)((bits8.s3 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    elem = (char)((bits8.s3 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    elem = (char)((bits8.s3 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    elem = (char)(bits8.s4 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    elem = (char)((bits8.s4 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    elem = (char)((bits8.s4 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    elem = (char)((bits8.s4 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    elem = (char)(bits8.s5 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    elem = (char)((bits8.s5 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    elem = (char)((bits8.s5 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    elem = (char)((bits8.s5 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    elem = (char)(bits8.s6 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    elem = (char)((bits8.s6 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    elem = (char)((bits8.s6 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    elem = (char)((bits8.s6 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+                                             \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    elem = (char)(bits8.s7 & 0x000000FF); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    elem = (char)((bits8.s7 & 0x0000FF00) >> 8); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    elem = (char)((bits8.s7 & 0x00FF0000) >> 16); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    elem = (char)((bits8.s7 & 0xFF000000) >> 24); \
+    total_sums += convert_int(elem) * scale * shared_y; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half  * src0_d,  // A scales
+        __read_only  image1d_buffer_t src1,    // B
+        ulong offset1,            // offset to B (0)
+        global float * dst,     // C
+        ulong offsetd,            // offset to C
+        int ne00,               // K
+        int ne01,               // M
+        int ne02,               // 1
+        int ne10,               // K
+        int ne12,               // 1
+        int ne0,                // M
+        int ne1,                // N
+        int r2,                 // 1
+        int r3)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid    = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A = M;
+    uint BLOCK_STRIDE_A = 8 * M;   // 32 / 4 = 8
+
+    __private uint8     regA;
+    __private half      regS;
+    __private float8    regB;
+
+    __private float totalSum = (float)(0.0f);
+
+    // loop along K in block granularity, skip 4 blocks every iter
+    #pragma unroll 1 /* tell compiler not to unroll */
+    for (uint k = groupId; k < (K / QK8_0); k += N_SIMDGROUP) {
+        regS = src0_d[gid + k * LINE_STRIDE_A]; // each fiber loads scale of one rows
+        // first 4 fibers in each wave load 8 B values to its private scope
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load weights for one block in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA.s4 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s5 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, regS, regB);
+    }
+
+    // reduction in local memory, assumes #wave=4
+    __local float reduceLM[SIMDGROUP_WIDTH * 3];
+    if (groupId == 1) reduceLM[SIMDGROUP_WIDTH * 0 + slid] = totalSum;
+    if (groupId == 2) reduceLM[SIMDGROUP_WIDTH * 1 + slid] = totalSum;
+    if (groupId == 3) reduceLM[SIMDGROUP_WIDTH * 2 + slid] = totalSum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 0 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 1 + slid];
+    if (groupId == 0) totalSum += reduceLM[SIMDGROUP_WIDTH * 2 + slid];
+
+    // 1 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        dst[gid] = totalSum;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl
new file mode 100644
index 0000000..c2962ed
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/get_rows.cl
@@ -0,0 +1,187 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+#define QK4_0                   32
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+
+//------------------------------------------------------------------------------
+// dequantize_q4_0_f32, dequantize_q4_0_f16
+//------------------------------------------------------------------------------
+void dequantize_q4_0_f32(global struct block_q4_0 * xb, short il, float16 * reg) {
+    global ushort * qs = ((global ushort *)xb + 1);
+    float d1 = il ? (xb->d / 16.h) : xb->d;
+    float d2 = d1 / 256.f;
+    float md = -8.h * xb->d;
+    ushort mask0 = il ? 0x00F0 : 0x000F;
+    ushort mask1 = mask0 << 8;
+
+    reg->s0 = d1 * (qs[0] & mask0) + md;
+    reg->s1 = d2 * (qs[0] & mask1) + md;
+
+    reg->s2 = d1 * (qs[1] & mask0) + md;
+    reg->s3 = d2 * (qs[1] & mask1) + md;
+
+    reg->s4 = d1 * (qs[2] & mask0) + md;
+    reg->s5 = d2 * (qs[2] & mask1) + md;
+
+    reg->s6 = d1 * (qs[3] & mask0) + md;
+    reg->s7 = d2 * (qs[3] & mask1) + md;
+
+    reg->s8 = d1 * (qs[4] & mask0) + md;
+    reg->s9 = d2 * (qs[4] & mask1) + md;
+
+    reg->sa = d1 * (qs[5] & mask0) + md;
+    reg->sb = d2 * (qs[5] & mask1) + md;
+
+    reg->sc = d1 * (qs[6] & mask0) + md;
+    reg->sd = d2 * (qs[6] & mask1) + md;
+
+    reg->se = d1 * (qs[7] & mask0) + md;
+    reg->sf = d2 * (qs[7] & mask1) + md;
+}
+
+
+//------------------------------------------------------------------------------
+// get_rows
+//------------------------------------------------------------------------------
+kernel void kernel_get_rows_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
+
+    int r = ((global int *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+    int i03 = i12;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        if (ind >= ne00) {
+            return;
+        }
+        ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] =
+            ((global float *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind];
+    }
+}
+
+kernel void kernel_get_rows_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
+
+    int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+    int i03 = i12;
+
+    for (int ind = get_local_id(0); ind < ne00; ind += get_local_size(0)) {
+        if (ind >= ne00) {
+            return;
+        }
+        ((global float *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1))[ind] =
+            ((global half *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03))[ind];
+    }
+}
+
+kernel void kernel_get_rows_q4_0(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    const int NL = 2;
+
+    int i10 = get_group_id(0);
+    int i11 = get_group_id(1);
+    int i12 = get_group_id(2);
+
+    int r = ((global int32_t *) ((global char *) src1 + i12*nb12 + i11*nb11 + i10*nb10))[0];
+
+    int i02 = i11;
+    int i03 = i12;
+
+    for (int ind = get_local_id(0); ind < ne00/16; ind += get_local_size(0)) {
+        float16 temp;
+        if (ind >= ne00) {
+            return;
+        }
+        dequantize_q4_0_f32(
+            ((global struct block_q4_0 *) ((global char *) src0 + r*nb01 + i02*nb02 + i03*nb03)) + ind/NL, ind%NL, &temp);
+        *(((global float16 *) ((global char *) dst + i12*nb3 + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl
new file mode 100644
index 0000000..059a4bb
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/glu.cl
@@ -0,0 +1,378 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define GELU_COEF_A     0.044715f
+#define GELU_QUICK_COEF -1.702f
+#define SQRT_2_OVER_PI  0.79788456080286535587989211986876f
+#define SQRT_2_INV      0.70710678118654752440084436210484f
+
+//------------------------------------------------------------------------------
+// geglu
+//------------------------------------------------------------------------------
+kernel void kernel_geglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+kernel void kernel_geglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu = 0.5f*x0*(1.0f + tanh(SQRT_2_OVER_PI*x0*(1.0f + GELU_COEF_A*x0*x0)));
+
+        dst_row[i0] = gelu*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// reglu
+//------------------------------------------------------------------------------
+kernel void kernel_reglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+kernel void kernel_reglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        dst_row[i0] = x0*x1*(x0 > 0.0f);
+    }
+}
+
+//------------------------------------------------------------------------------
+// swiglu
+//------------------------------------------------------------------------------
+kernel void kernel_swiglu(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+kernel void kernel_swiglu_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half silu = x0 / (1.0f + exp(-x0));
+
+        dst_row[i0] = silu*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// swiglu_oai
+//------------------------------------------------------------------------------
+kernel void kernel_swiglu_oai(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb01,
+    ulong         nb11,
+    int           ne0,
+    ulong         nb1,
+    int           ne00_off,
+    int           ne10_off,
+    float         limit,
+    float         alpha
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        float x0 = src0_row[i0];
+        float x1 = src1_row[i0];
+
+        x0 = min(x0, limit);
+        x1 = max(min(x1, limit), -limit);
+
+        float out_glu = x0 / (1.0f + exp(-x0 * alpha));
+        out_glu = out_glu * (1.0f + x1);
+
+        dst_row[i0] = out_glu;
+    }
+}
+
+//------------------------------------------------------------------------------
+// geglu_erf
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_erf(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+kernel void kernel_geglu_erf_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_erf = 0.5f*x0*(1.0f + erf(x0*SQRT_2_INV));
+
+        dst_row[i0] = gelu_erf*x1;
+    }
+}
+
+//------------------------------------------------------------------------------
+// geglu_quick
+//------------------------------------------------------------------------------
+kernel void kernel_geglu_quick(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global float * src0_row = (global float *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global float * src1_row = (global float *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global float * dst_row  = (global float *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const float x0 = src0_row[i0];
+        const float x1 = src1_row[i0];
+
+        const float gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
+
+kernel void kernel_geglu_quick_f16(
+    global char * src0,
+    ulong  offset0,
+    global char * src1,
+    ulong  offset1,
+    global char * dst,
+    ulong  offsetd,
+    ulong nb01,
+    ulong nb11,
+    int ne0,
+    ulong nb1,
+    int ne00_off,
+    int ne10_off
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    global half * src0_row = (global half *) ((global char *) src0 + get_group_id(0)*nb01) + ne00_off;
+    global half * src1_row = (global half *) ((global char *) src1 + get_group_id(0)*nb11) + ne10_off;
+    global half * dst_row  = (global half *) ((global char *) dst  + get_group_id(0)*nb1);
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const half x0 = src0_row[i0];
+        const half x1 = src1_row[i0];
+
+        const half gelu_quick = x0*(1.0f/(1.0f + exp(GELU_QUICK_COEF*x0)));
+
+        dst_row[i0] = gelu_quick*x1;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl
new file mode 100644
index 0000000..8e4fa0e
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/group_norm.cl
@@ -0,0 +1,121 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Workgroup must be a subgroup
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_group_norm(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne,
+        int group_size,
+        float eps
+) {
+    src0 = (global float  *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int start = get_group_id(0) * group_size;
+    int end   = start + group_size;
+
+    start += get_local_id(0);
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f;
+
+    for (int j = start; j < end; j += get_local_size(0)) {
+        tmp += src0[j];
+    }
+
+    tmp = sub_group_reduce_add(tmp);
+
+    const float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += get_local_size(0)) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = sub_group_reduce_add(tmp);
+
+    const float variance = tmp / group_size;
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int j = start; j < end; j += get_local_size(0)) {
+        dst[j] *= scale;
+    }
+}
+
+//------------------------------------------------------------------------------
+// group_norm_mul_add
+//------------------------------------------------------------------------------
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_group_norm_mul_add(
+        global float * src0, ulong offset0,
+        global float * src1, ulong offset1,
+        global float * src2, ulong offset2,
+        global float * dst, ulong offsetd,
+        int ne,
+        int group_size,
+        float eps
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global float *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int start = get_group_id(0) * group_size;
+    int end = start + group_size;
+    if (end > ne) {
+        end = ne;
+    }
+
+    float sum = 0.0f;
+    float sum_sq = 0.0f;
+
+    for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) {
+        float val = src0[j];
+        sum += val;
+        sum_sq += val*val;
+    }
+
+    sum = sub_group_reduce_add(sum);
+    sum_sq = sub_group_reduce_add(sum_sq);
+
+    const float mean = sum / group_size;
+    const float var = sum_sq / group_size - mean * mean;
+    const float scale = rsqrt(var + eps);
+
+    for (int j = start + get_local_id(0); j < end; j += get_local_size(0)) {
+        dst[j] = ((src0[j] - mean) * scale) * src1[j] + src2[j];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl
new file mode 100644
index 0000000..cf6cdaa
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f16.cl
@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f16(
+        global float * src1,
+        ulong offset1,
+        global half  * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    long  ksize = OW * KH;
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl
new file mode 100644
index 0000000..1ecdb23
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/im2col_f32.cl
@@ -0,0 +1,57 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_im2col_f32(
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        ulong batch_offset,
+        ulong delta_offset,
+        long IW,
+        long IH,
+        long IC,
+        long OW,
+        long OH,
+        long KW,
+        long KH,
+        long pelements,
+        long CHW,
+        int  s0,
+        int  s1,
+        int  p0,
+        int  p1,
+        int  d0,
+        int  d1
+) {
+    long i = get_global_id(0);
+    if (i >= pelements) {
+        return;
+    }
+
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    long  ksize = OW * KH;
+    long  kx = i / ksize;
+    long  kd = kx * ksize;
+    long  ky = (i - kd) / OW;
+    long  ix = i % OW;
+
+    long  oh = get_group_id(1);
+    long  batch = get_group_id(2) / IC;
+    long  ic = get_group_id(2) % IC;
+
+    long iiw = ix * s0 + kx * d0 - p0;
+    long iih = oh * s1 + ky * d1 - p1;
+
+    long offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        long offset_src = ic * delta_offset + batch * batch_offset;
+        dst[offset_dst] = src1[offset_src + iih * IW + iiw];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl
new file mode 100644
index 0000000..5c3e8bc
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mean.cl
@@ -0,0 +1,39 @@
+
+kernel void kernel_mean_f32(
+    global float *  src0,
+    ulong           offset0,
+    global float *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int i3 = get_global_id(2);
+    int i2 = get_global_id(1);
+    int i1 = get_global_id(0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) ((global char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum / ne00;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl
new file mode 100644
index 0000000..b12a592
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul.cl
@@ -0,0 +1,152 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// mul
+//------------------------------------------------------------------------------
+kernel void kernel_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) * *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_mul_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}
+
+kernel void kernel_mul_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) * *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] * src1[idx1];
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl
new file mode 100644
index 0000000..ecb577b
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_Ab_Bi_8x4.cl
@@ -0,0 +1,139 @@
+// src0_q, src0_d, src1 are transposed as a preprocessing step
+// 4-bit weights are transposed in groups of 4 (unsigned short int)
+// consider weights originally "next to each other", now "on top of each other"
+// each fiber computes a 8x4 tile of output elements
+// using unshuffled weights
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_mul_mat_Ab_Bi_8x4(
+        global const ushort * src0_q,       // quantized A
+        global const half  * src0_d,        // A scales
+        __read_only image1d_buffer_t src1,  // B (1d image)
+        global float * dst,                 // C
+        int m,                              // M
+        int n,                              // N with padding
+        int k,                              // K
+        int n_no_padding                    // N without padding
+) {
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0; // 8x4 output elements
+    half8 B; // registers for activations
+    half4 dequantized_weights; // registers for dequantized weights
+    __global const ushort* weight_ptr = src0_q + gx_2; // pointer for weights
+    __global const half* scale_ptr = src0_d + gx_2; // pointer for scales
+
+    for(int i=0; i<k; i+=4){ //loop through K dimension
+
+        B.s0123 = read_imageh(src1, gy*2 + (i)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i)*(n_4)+1);
+
+        // keep (i/4) and (i/32) in parenthesis, rounds down
+        // load 4 consecutive groups of 4 weights
+        ushort4 bits4 = vload4(0, weight_ptr + (i/4)*(m)); // (i/4) because weights grouped in 4s
+
+        // load 4 consecutive scales
+        half4 scale = vload4(0, scale_ptr + (i/32)*(m));// (i/32) because 1 scale per 32 elements
+
+        // j=0
+        dequantized_weights.s0 = ((bits4.s0 & (0x000F)) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = ((bits4.s1 & (0x000F)) - 8) * scale.s1;
+        dequantized_weights.s2 = ((bits4.s2 & (0x000F)) - 8) * scale.s2;
+        dequantized_weights.s3 = ((bits4.s3 & (0x000F)) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=1
+        B.s0123 = read_imageh(src1, gy*2 + (i+1)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+1)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x00F0)) >> 4) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x00F0)) >> 4) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x00F0)) >> 4) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x00F0)) >> 4) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; //vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=2
+        B.s0123 = read_imageh(src1, gy*2 + (i+2)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+2)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0x0F00)) >> 8) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0x0F00)) >> 8) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0x0F00)) >> 8) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0x0F00)) >> 8) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+
+        // j=3
+        B.s0123 = read_imageh(src1, gy*2 + (i+3)*(n_4));
+        B.s4567 = read_imageh(src1, gy*2 + (i+3)*(n_4)+1);
+        dequantized_weights.s0 = (((bits4.s0 & (0xF000)) >> 12) - 8) * scale.s0; // dequantize a row of the 16 weights
+        dequantized_weights.s1 = (((bits4.s1 & (0xF000)) >> 12) - 8) * scale.s1;
+        dequantized_weights.s2 = (((bits4.s2 & (0xF000)) >> 12) - 8) * scale.s2;
+        dequantized_weights.s3 = (((bits4.s3 & (0xF000)) >> 12) - 8) * scale.s3;
+        c0 += B * dequantized_weights.s0; // vector-scalar multiplication to accumulate
+        c1 += B * dequantized_weights.s1;
+        c2 += B * dequantized_weights.s2;
+        c3 += B * dequantized_weights.s3;
+    }
+
+    int idx = (gy<<3)*m + (gx<<2); // vectorized store 16 elements
+
+    // conditional check if store is to a valid location. Required when N is not a multiple of 8
+    // if statements allow registers to be reused for each store
+    // provides a performance boost due to reduced register footprint, which increases number of concurrent waves
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl
new file mode 100644
index 0000000..73a8884
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mat_f16_f32.cl
@@ -0,0 +1,130 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#if defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#else
+#define REQD_SUBGROUP_SIZE_128
+#endif
+
+#define OPWM 64
+#define OPWN 64
+#define CPWK 8
+#define OPTM 4
+#define OPTN 8
+
+#define WG_M (OPWM / OPTM)
+#define WG_N (OPWN / OPTN)
+#define VEC_K (CPWK / 4)
+
+REQD_SUBGROUP_SIZE_128
+__kernel void mul_mat_f16_f32(
+    const int M, const int N, const int K,
+    __global const void* A_void, ulong A_offset,
+    __global const void* B_void, ulong B_offset,
+    __global       void* C_void, ulong C_offset) {
+
+    __global const half*  A = (__global const half* )((__global const char*)A_void + A_offset);
+    __global const float* B = (__global const float*)((__global const char*)B_void + B_offset);
+    __global       float* C = (__global       float*)((__global       char*)C_void + C_offset);
+
+    const int lidm = get_local_id(0);
+    const int lidn = get_local_id(1);
+    const int lid = lidn * WG_M + lidm;
+
+    const int offsetM = get_group_id(0) * OPWM;
+    const int offsetN = get_group_id(1) * OPWN;
+
+    __local half4  Alocal[OPWM][VEC_K];
+    __local float4 Blocal[OPWN][VEC_K];
+
+    float sum[OPTM][OPTN];
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        for (int wn = 0; wn < OPTN; wn++) {
+            sum[wm][wn] = 0.0f;
+        }
+    }
+
+    const int numTiles = (K + CPWK - 1) / CPWK;
+
+    const int load_row_a = lid % OPWM;
+    const int load_vec_k_a = lid / OPWM;
+    const int global_row_a = offsetM + load_row_a;
+
+    const int load_row_b = lid % OPWN;
+    const int load_vec_k_b = lid / OPWN;
+    const int global_row_b = offsetN + load_row_b;
+
+    for (int t = 0; t < numTiles; t++) {
+        const int k_start = t * CPWK;
+        const int k_vec_start_a = k_start + load_vec_k_a * 4;
+        const int k_vec_start_b = k_start + load_vec_k_b * 4;
+
+        if (global_row_a < M && k_vec_start_a < K) {
+            if (k_vec_start_a + 3 < K) {
+                Alocal[load_row_a][load_vec_k_a] = vload4(0, A + global_row_a * K + k_vec_start_a);
+            } else {
+                half4 tempA = (half4)(0.0h);
+                if (k_vec_start_a < K) tempA.s0 = A[global_row_a * K + k_vec_start_a];
+                if (k_vec_start_a + 1 < K) tempA.s1 = A[global_row_a * K + k_vec_start_a + 1];
+                if (k_vec_start_a + 2 < K) tempA.s2 = A[global_row_a * K + k_vec_start_a + 2];
+                Alocal[load_row_a][load_vec_k_a] = tempA;
+            }
+        } else {
+            Alocal[load_row_a][load_vec_k_a] = (half4)(0.0h);
+        }
+
+        if (global_row_b < N && k_vec_start_b < K) {
+            if (k_vec_start_b + 3 < K) {
+                Blocal[load_row_b][load_vec_k_b] = vload4(0, B + global_row_b * K + k_vec_start_b);
+            } else {
+                float4 tempB = (float4)(0.0f);
+                if (k_vec_start_b < K) tempB.s0 = B[global_row_b * K + k_vec_start_b];
+                if (k_vec_start_b + 1 < K) tempB.s1 = B[global_row_b * K + k_vec_start_b + 1];
+                if (k_vec_start_b + 2 < K) tempB.s2 = B[global_row_b * K + k_vec_start_b + 2];
+                Blocal[load_row_b][load_vec_k_b] = tempB;
+            }
+        } else {
+            Blocal[load_row_b][load_vec_k_b] = (float4)(0.0f);
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        #pragma unroll
+        for (int k_vec = 0; k_vec < VEC_K; k_vec++) {
+            float4 a_fvecs[OPTM];
+            int current_row_a = lidm;
+            for (int wm = 0; wm < OPTM; wm++) {
+                a_fvecs[wm] = convert_float4(Alocal[current_row_a][k_vec]);
+                current_row_a += WG_M;
+            }
+
+            float4 b_fvecs[OPTN];
+            int current_row_b = lidn;
+            for (int wn = 0; wn < OPTN; wn++) {
+                b_fvecs[wn] = Blocal[current_row_b][k_vec];
+                current_row_b += WG_N;
+            }
+
+            for (int wm = 0; wm < OPTM; wm++) {
+                for (int wn = 0; wn < OPTN; wn++) {
+                    sum[wm][wn] += dot(a_fvecs[wm], b_fvecs[wn]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    for (int wm = 0; wm < OPTM; wm++) {
+        int globalRow = offsetM + lidm + wm * WG_M;
+        if (globalRow < M) {
+            for (int wn = 0; wn < OPTN; wn++) {
+                int globalCol = offsetN + lidn + wn * WG_N;
+                if (globalCol < N) {
+                    C[globalCol * M + globalRow] = sum[wm][wn];
+                }
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl
new file mode 100644
index 0000000..ac0274b
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_kq_kqv.cl
@@ -0,0 +1,273 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#define LM_FIRST_256B   0
+#define LM_SECOND_256B  64
+#define LM_THIRD_256B   128
+#define LM_FOURTH_256B  192
+
+
+inline float16 mm_load_a(
+    image1d_buffer_t matrix_A,
+    uint subMatrixAStartInElements,
+    int nb01,
+    int line_stride_matrix_A_in_bytes
+) {
+    __private float8 regA;
+    size_t sub_block_id_m = get_local_id(0);
+
+#ifdef KQV
+    uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * nb01/4);
+#else // KQ
+    uint a_texCoord = subMatrixAStartInElements/2 + (sub_block_id_m * line_stride_matrix_A_in_bytes/4);
+#endif
+
+    regA.s0123  = read_imagef(matrix_A, a_texCoord/4);
+    regA.s4567  = read_imagef(matrix_A, (a_texCoord+4)/4);
+
+    return convert_float16(as_half16(regA));
+}
+
+inline float4 alu_32(
+    float16 regA,
+    __local float4* matrix_B_vec
+) {
+
+    __private float4 rC = 0;
+    int i = get_sub_group_id() * 64;
+
+    rC += regA.s0  * matrix_B_vec[i];
+    rC += regA.s1  * matrix_B_vec[i + 16];
+    rC += regA.s4  * matrix_B_vec[i + 1];
+    rC += regA.s5  * matrix_B_vec[i + 17];
+    rC += regA.s8  * matrix_B_vec[i + 2];
+    rC += regA.s9  * matrix_B_vec[i + 18];
+    rC += regA.sc  * matrix_B_vec[i + 3];
+    rC += regA.sd  * matrix_B_vec[i + 19];
+
+    i += 32;
+
+    rC += regA.s2  * matrix_B_vec[i];
+     rC += regA.s3  * matrix_B_vec[i + 16];
+    rC += regA.s6  * matrix_B_vec[i + 1];
+    rC += regA.s7  * matrix_B_vec[i + 17];
+    rC += regA.sa  * matrix_B_vec[i + 2];
+    rC += regA.sb  * matrix_B_vec[i + 18];
+    rC += regA.se  * matrix_B_vec[i + 3];
+    rC += regA.sf  * matrix_B_vec[i + 19];
+
+    return rC;
+}
+
+inline float16 alu_16(
+    float16 regA,
+    __local float* matrix_B_local
+) {
+    float16 out;
+    __local float4* matrix_B_vec = (__local float4*)matrix_B_local;
+
+    out.s0123 = alu_32(regA, matrix_B_vec);
+    out.s4567 = alu_32(regA, matrix_B_vec + 4);
+    out.s89ab = alu_32(regA, matrix_B_vec + 8);
+    out.scdef = alu_32(regA, matrix_B_vec + 12);
+
+    return out;
+}
+
+inline void mm_mad(
+    __local float* matrix_B_local,
+    float16 regA,
+    float8 regB,
+    uint b_localOffsetInWords,
+    float16* regC0_ptr,
+    float16* regC1_ptr
+) {
+    int offset = b_localOffsetInWords + get_sub_group_id() * 256;
+
+    matrix_B_local[offset + LM_FIRST_256B] = regB.s0;
+    matrix_B_local[offset + LM_SECOND_256B] = regB.s1;
+    matrix_B_local[offset + LM_THIRD_256B] = regB.s2;
+    matrix_B_local[offset + LM_FOURTH_256B] = regB.s3;
+
+    float16 add0 = alu_16(regA, matrix_B_local);
+    *regC0_ptr += add0;
+
+    matrix_B_local[offset + LM_FIRST_256B] = regB.s4;
+    matrix_B_local[offset + LM_SECOND_256B] = regB.s5;
+    matrix_B_local[offset + LM_THIRD_256B] = regB.s6;
+    matrix_B_local[offset + LM_FOURTH_256B] = regB.s7;
+
+    float16 add1 = alu_16(regA, matrix_B_local);
+    *regC1_ptr += add1;
+}
+
+inline void mm_store_c_N(
+    __write_only image1d_buffer_t matrix_C,
+    float16 regC0,
+    float16 regC1,
+    uint subMatrixCStartInElements,
+    int line_stride_matrix_C_in_bytes,
+    int mask
+) {
+    size_t sub_block_id_m = get_local_id(0);
+
+    uint strideInWords     = line_stride_matrix_C_in_bytes/4;
+    uint c_coordInWords_0  = (subMatrixCStartInElements + sub_block_id_m);
+
+    uint c_coordInWords_1  = c_coordInWords_0 + 1  * strideInWords;
+    uint c_coordInWords_2  = c_coordInWords_0 + 2  * strideInWords;
+    uint c_coordInWords_3  = c_coordInWords_0 + 3  * strideInWords;
+    uint c_coordInWords_4  = c_coordInWords_0 + 4  * strideInWords;
+    uint c_coordInWords_5  = c_coordInWords_0 + 5  * strideInWords;
+    uint c_coordInWords_6  = c_coordInWords_0 + 6  * strideInWords;
+    uint c_coordInWords_7  = c_coordInWords_0 + 7  * strideInWords;
+    uint c_coordInWords_8  = c_coordInWords_0 + 8  * strideInWords;
+    uint c_coordInWords_9  = c_coordInWords_0 + 9  * strideInWords;
+    uint c_coordInWords_10 = c_coordInWords_0 + 10 * strideInWords;
+    uint c_coordInWords_11 = c_coordInWords_0 + 11 * strideInWords;
+    uint c_coordInWords_12 = c_coordInWords_0 + 12 * strideInWords;
+    uint c_coordInWords_13 = c_coordInWords_0 + 13 * strideInWords;
+    uint c_coordInWords_14 = c_coordInWords_0 + 14 * strideInWords;
+    uint c_coordInWords_15 = c_coordInWords_0 + 15 * strideInWords;
+    uint c_coordInWords_16 = c_coordInWords_0 + 16 * strideInWords;
+    uint c_coordInWords_17 = c_coordInWords_0 + 17 * strideInWords;
+    uint c_coordInWords_18 = c_coordInWords_0 + 18 * strideInWords;
+    uint c_coordInWords_19 = c_coordInWords_0 + 19 * strideInWords;
+    uint c_coordInWords_20 = c_coordInWords_0 + 20 * strideInWords;
+    uint c_coordInWords_21 = c_coordInWords_0 + 21 * strideInWords;
+    uint c_coordInWords_22 = c_coordInWords_0 + 22 * strideInWords;
+    uint c_coordInWords_23 = c_coordInWords_0 + 23 * strideInWords;
+    uint c_coordInWords_24 = c_coordInWords_0 + 24 * strideInWords;
+    uint c_coordInWords_25 = c_coordInWords_0 + 25 * strideInWords;
+    uint c_coordInWords_26 = c_coordInWords_0 + 26 * strideInWords;
+    uint c_coordInWords_27 = c_coordInWords_0 + 27 * strideInWords;
+    uint c_coordInWords_28 = c_coordInWords_0 + 28 * strideInWords;
+    uint c_coordInWords_29 = c_coordInWords_0 + 29 * strideInWords;
+    uint c_coordInWords_30 = c_coordInWords_0 + 30 * strideInWords;
+    uint c_coordInWords_31 = c_coordInWords_0 + 31 * strideInWords;
+
+    if (mask > 0)  { write_imagef(matrix_C, c_coordInWords_0, regC0.s0);  }
+    if (mask > 1)  { write_imagef(matrix_C, c_coordInWords_1, regC0.s1);  }
+    if (mask > 2)  { write_imagef(matrix_C, c_coordInWords_2, regC0.s2);  }
+    if (mask > 3)  { write_imagef(matrix_C, c_coordInWords_3, regC0.s3);  }
+    if (mask > 4)  { write_imagef(matrix_C, c_coordInWords_4, regC0.s4);  }
+    if (mask > 5)  { write_imagef(matrix_C, c_coordInWords_5, regC0.s5);  }
+    if (mask > 6)  { write_imagef(matrix_C, c_coordInWords_6, regC0.s6);  }
+    if (mask > 7)  { write_imagef(matrix_C, c_coordInWords_7, regC0.s7);  }
+    if (mask > 8)  { write_imagef(matrix_C, c_coordInWords_8, regC0.s8);  }
+    if (mask > 9)  { write_imagef(matrix_C, c_coordInWords_9, regC0.s9);  }
+    if (mask > 10) { write_imagef(matrix_C, c_coordInWords_10, regC0.sa); }
+    if (mask > 11) { write_imagef(matrix_C, c_coordInWords_11, regC0.sb); }
+    if (mask > 12) { write_imagef(matrix_C, c_coordInWords_12, regC0.sc); }
+    if (mask > 13) { write_imagef(matrix_C, c_coordInWords_13, regC0.sd); }
+    if (mask > 14) { write_imagef(matrix_C, c_coordInWords_14, regC0.se); }
+    if (mask > 15) { write_imagef(matrix_C, c_coordInWords_15, regC0.sf); }
+    if (mask > 16) { write_imagef(matrix_C, c_coordInWords_16, regC1.s0); }
+    if (mask > 17) { write_imagef(matrix_C, c_coordInWords_17, regC1.s1); }
+    if (mask > 18) { write_imagef(matrix_C, c_coordInWords_18, regC1.s2); }
+    if (mask > 19) { write_imagef(matrix_C, c_coordInWords_19, regC1.s3); }
+    if (mask > 20) { write_imagef(matrix_C, c_coordInWords_20, regC1.s4); }
+    if (mask > 21) { write_imagef(matrix_C, c_coordInWords_21, regC1.s5); }
+    if (mask > 22) { write_imagef(matrix_C, c_coordInWords_22, regC1.s6); }
+    if (mask > 23) { write_imagef(matrix_C, c_coordInWords_23, regC1.s7); }
+    if (mask > 24) { write_imagef(matrix_C, c_coordInWords_24, regC1.s8); }
+    if (mask > 25) { write_imagef(matrix_C, c_coordInWords_25, regC1.s9); }
+    if (mask > 26) { write_imagef(matrix_C, c_coordInWords_26, regC1.sa); }
+    if (mask > 27) { write_imagef(matrix_C, c_coordInWords_27, regC1.sb); }
+    if (mask > 28) { write_imagef(matrix_C, c_coordInWords_28, regC1.sc); }
+    if (mask > 29) { write_imagef(matrix_C, c_coordInWords_29, regC1.sd); }
+    if (mask > 30) { write_imagef(matrix_C, c_coordInWords_30, regC1.se); }
+    if (mask > 31) { write_imagef(matrix_C, c_coordInWords_31, regC1.sf); }
+}
+
+#define TILESIZE_K 16
+#define TILESIZE_M 64
+#define TILESIZE_N 32
+#ifdef KQV
+__kernel void mul_mm_f16_f32_kqv(
+#else
+__kernel void mul_mm_f16_f32_kq(
+#endif
+        __read_only  image1d_buffer_t matrix_A,
+        int offset0,
+        __global float* matrix_B,
+        int offset1,
+        __write_only image1d_buffer_t matrix_C,
+        int offsetd,
+        int M, int K, int N,
+        int D_A,
+        int D_B,
+        int nb01
+) {
+
+    uint block_id_m = get_global_id(1);
+    uint block_id_n = get_global_id(2) % ((N+TILESIZE_N-1)/TILESIZE_N);
+    uint block_id_d = get_global_id(2) / ((N+TILESIZE_N-1)/TILESIZE_N);
+
+    __private float16  regA;
+    __private float8   regB;
+    __private float16 regC0;
+    __private float16 regC1;
+
+    const uint col   = block_id_m * TILESIZE_M;
+    const uint row   = block_id_n * TILESIZE_N;
+    const uint depth_A = block_id_d / (D_B/D_A);
+    const uint depth_B = block_id_d;
+
+#ifdef KQV
+    int line_stride_matrix_A_in_bytes = nb01 * M;
+    int line_stride_matrix_B_in_bytes = K * N * 4;
+#else
+    int line_stride_matrix_A_in_bytes = K * D_A * 2;
+    int line_stride_matrix_B_in_bytes = K * D_B * 4;
+#endif
+
+    int line_stride_matrix_C_in_bytes = M * 4;
+
+    const uint strideAinElements = line_stride_matrix_A_in_bytes / 2;
+    const uint strideBinElements = line_stride_matrix_B_in_bytes / 4;
+
+    size_t sub_block_id_m = get_local_id(0);
+
+    uint b_localOffsetInWords = (sub_block_id_m/16)*16
+                           + ((((sub_block_id_m)>>0)&1)<<2)
+                           + ((((sub_block_id_m)>>1)&1)<<3)
+                           + ((((sub_block_id_m)>>2)&1)<<0)
+                           + ((((sub_block_id_m)>>3)&1)<<1);
+
+    uint2 b_globalOffsetInWords_xy = {((sub_block_id_m%4)*4), (sub_block_id_m>>2)};
+    uint b_globalOffsetInWords00, b_globalOffsetInWords16;
+#ifdef KQV
+    b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*K;
+    b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * K);
+    uint subMatrixAStartInElements = depth_A * strideAinElements + col * nb01 / 2;
+    uint subMatrixBStartInElements = depth_B * strideBinElements + row * K;
+#else
+    b_globalOffsetInWords00 = b_globalOffsetInWords_xy.x + b_globalOffsetInWords_xy.y*line_stride_matrix_B_in_bytes/4;
+    b_globalOffsetInWords16 = b_globalOffsetInWords00 + (16 * line_stride_matrix_B_in_bytes/4);
+    uint subMatrixAStartInElements = col * strideAinElements + depth_A * K;
+    uint subMatrixBStartInElements = row * strideBinElements + depth_B * K;
+#endif
+
+    __local float matrix_B_local[1024];
+
+    for (uint step=0; step < K; step+=TILESIZE_K) {
+        size_t sub_block_id_m = get_local_id(0);
+        regA = mm_load_a(matrix_A, subMatrixAStartInElements, nb01, line_stride_matrix_A_in_bytes);
+
+        uint b_coordInWords00 = subMatrixBStartInElements + b_globalOffsetInWords00;
+        uint b_coordInWords16 = subMatrixBStartInElements + b_globalOffsetInWords16;
+
+        regB.s0123 = vload4(b_coordInWords00/4, matrix_B);
+        regB.s4567 = vload4(b_coordInWords16/4, matrix_B);
+
+        mm_mad(matrix_B_local, regA, regB, b_localOffsetInWords, &regC0, &regC1);
+
+        subMatrixAStartInElements += TILESIZE_K;
+        subMatrixBStartInElements += TILESIZE_K;
+    }
+
+    uint subMatrixCStartInElements = depth_B * N * M + row * M + col;
+    mm_store_c_N(matrix_C, regC0, regC1, subMatrixCStartInElements, line_stride_matrix_C_in_bytes, (N-block_id_n*32));
+}
+
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl
new file mode 100644
index 0000000..6982f8f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f16_f32_l4_lm.cl
@@ -0,0 +1,146 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f16_f32_l4_lm(
+    global half4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local half  buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    half  cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0h;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0h;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0h;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0h;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0h;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0h;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0h;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0h;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(convert_float(cache_a[cr]), cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl
new file mode 100644
index 0000000..d7d5ba6
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_f32_f32_l4_lm.cl
@@ -0,0 +1,147 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 16
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_f32_f32_l4_lm(
+    global float4 * src0,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                const int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = src0[idx].s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = src0[idx].s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = src0[idx].s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = src0[idx].s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                const int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl
new file mode 100644
index 0000000..3602c92
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q6_k_f32_l4_lm.cl
@@ -0,0 +1,158 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 2
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q6_k_f32_l4_lm(
+    global uchar * src0_ql,
+    global uchar * src0_qh,
+    global char  * src0_s,
+    global half  * src0_d,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+
+                int ib = idx / 128;                  // 2 values per idx
+                int iqs = idx % 128;                 // 0..127
+
+                int n = iqs / 64;                    // 0,1
+                int b = (iqs % 64) / 32;             // 0,1
+                int is_b = (iqs % 16) / 8;           // 0,1
+                int qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+                int is = 8 * n + qhshift + is_b;     // 0..15
+                int qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+                int qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+                float dscale = (float)src0_d[ib] * (float)src0_s[ib*16 + is];
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 0] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 0] >> qhshift) & 3) << 4)) - 32);
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = dscale * convert_float(convert_char(((src0_ql[128*ib + qsi + 1] >> (b * 4)) & 0xF) | (((src0_qh[64*ib + qhi + 1] >> qhshift) & 3) << 4)) - 32);
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl
new file mode 100644
index 0000000..51ce212
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_8x4.cl
@@ -0,0 +1,129 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+
+kernel void kernel_mul_mm_q8_0_f32_8x4(
+        global const uint * src0_q,
+        global const half  * src0_d,
+        __read_only image1d_buffer_t src1,
+        global float * dst,
+        int k,
+        int m,
+        int n,
+        int n_no_padding,
+        ulong offsetd
+) {
+
+    int m_4 = m >> 2;
+    int n_4 = n >> 2;
+
+    int gy   = get_global_id(0);
+    int gx   = get_global_id(1);
+    int gx_2 = gx << 2;
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 deq;
+
+    __global const uint* wptr = src0_q + gx_2;
+    __global const half* sptr = src0_d + gx_2;
+
+      for (int i = 0; i < k; i += 4) {
+        uint4 pack4 = vload4(0, wptr + (i / 4) * m);
+        half4 scale = vload4(0, sptr + (i / 32) * m);
+
+        char4 p0 = as_char4(pack4.s0);
+        char4 p1 = as_char4(pack4.s1);
+        char4 p2 = as_char4(pack4.s2);
+        char4 p3 = as_char4(pack4.s3);
+
+        // ------------------- j = 0 (k = i+0) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 0) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 0) * n_4 + 1);
+
+        half4 wj0 = convert_half4((char4)(p0.s0, p1.s0, p2.s0, p3.s0)) * scale;
+
+        c0 += B * wj0.s0;
+        c1 += B * wj0.s1;
+        c2 += B * wj0.s2;
+        c3 += B * wj0.s3;
+
+        // ------------------- j = 1 (k = i+1) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 1) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 1) * n_4 + 1);
+
+        half4 wj1 = convert_half4((char4)(p0.s1, p1.s1, p2.s1, p3.s1)) * scale;
+
+        c0 += B * wj1.s0;
+        c1 += B * wj1.s1;
+        c2 += B * wj1.s2;
+        c3 += B * wj1.s3;
+
+        // ------------------- j = 2 (k = i+2) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 2) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 2) * n_4 + 1);
+
+        half4 wj2 = convert_half4((char4)(p0.s2, p1.s2, p2.s2, p3.s2)) * scale;
+
+        c0 += B * wj2.s0;
+        c1 += B * wj2.s1;
+        c2 += B * wj2.s2;
+        c3 += B * wj2.s3;
+
+        // ------------------- j = 3 (k = i+3) -------------------
+        B.s0123 = read_imageh(src1, gy * 2 + (i + 3) * n_4);
+        B.s4567 = read_imageh(src1, gy * 2 + (i + 3) * n_4 + 1);
+
+        half4 wj3 = convert_half4((char4)(p0.s3, p1.s3, p2.s3, p3.s3)) * scale;
+
+        c0 += B * wj3.s0;
+        c1 += B * wj3.s1;
+        c2 += B * wj3.s2;
+        c3 += B * wj3.s3;
+    }
+
+    int idx = (gy << 3) * m + (gx << 2);
+
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if(idx+3 < m*n_no_padding){
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl
new file mode 100644
index 0000000..147b66f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mm_q8_0_f32_l4_lm.cl
@@ -0,0 +1,154 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#define TM 4
+#define TN 8
+
+kernel void kernel_mul_mm_q8_0_f32_l4_lm(
+    global char4  * src0_q,
+    global half   * src0_d,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = (batch_idx_a * batch_stride_a + ir * BM * stride_a) / LOAD_VEC_A;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a / LOAD_VEC_A + loadr_a;
+                int ib  = idx / 8;
+                int iqs = idx % 8;
+
+                float d = (float)src0_d[ib];
+                global char4 * qs = src0_q + ib*8 + iqs;
+                char4 q = *qs;
+                float4 v = convert_float4(q)*d;
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const int offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl
new file mode 100644
index 0000000..9393b54
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f16.cl
@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F16 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3)
+{
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F16;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half * y = (global half *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (half) x[i] * (half) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F16; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global half  * y  = (global half  *) (src1 + offset_src1);
+            global half4 * y4 = (global half4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (half) x4[i].s0 * y4[i].s0;
+                sumf += (half) x4[i].s1 * y4[i].s1;
+                sumf += (half) x4[i].s2 * y4[i].s2;
+                sumf += (half) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (half) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl
new file mode 100644
index 0000000..e52d3c6
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32.cl
@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F16_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F16_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half * x = (global half *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += convert_float(x[i]) * y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global half4 * x4 = (global half4 *)x;
+        for (int row = 0; row < N_F16_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += convert_float(x4[i].s0) * y4[i].s0;
+                sumf += convert_float(x4[i].s1) * y4[i].s1;
+                sumf += convert_float(x4[i].s2) * y4[i].s2;
+                sumf += convert_float(x4[i].s3) * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl
new file mode 100644
index 0000000..28d3021
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_1row.cl
@@ -0,0 +1,94 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_1row(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global half  * x = (global half  *) (src0 + offset_src0);
+    global float * y = (global float *) (src1 + offset_src1);
+
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        global half4  * x4 = (global half4  *) x;
+        global float4 * y4 = (global float4 *) y;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += (float) x4[i].s0 * y4[i].s0;
+            sumf += (float) x4[i].s1 * y4[i].s1;
+            sumf += (float) x4[i].s2 * y4[i].s2;
+            sumf += (float) x4[i].s3 * y4[i].s3;
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) {
+                all_sum += (float) x[i] * y[i];
+            }
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl
new file mode 100644
index 0000000..cdf8197
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f16_f32_l4.cl
@@ -0,0 +1,84 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Assumes row size (ne00) is a multiple of 4
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_l4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nrows = ne11;
+    int r0 = get_group_id(0);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global half4 * x4 = (global half4 *) (src0 + offset_src0);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+        global float4 * y4 = (global float4 *) (src1 + offset_src1);
+
+        float sumf = 0;
+        for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+            sumf += convert_float(x4[i].s0) * y4[i].s0;
+            sumf += convert_float(x4[i].s1) * y4[i].s1;
+            sumf += convert_float(x4[i].s2) * y4[i].s2;
+            sumf += convert_float(x4[i].s3) * y4[i].s3;
+        }
+
+        float all_sum = sub_group_reduce_add(sumf);
+        if (get_sub_group_local_id() == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl
new file mode 100644
index 0000000..ec71b87
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_f32_f32.cl
@@ -0,0 +1,118 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define N_F32_F32 4
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f32_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int r0 = get_group_id(0);
+    int rb = get_group_id(1)*N_F32_F32;
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global float * x = (global float *) (src0 + offset_src0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float * y = (global float *) (src1 + offset_src1);
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00; i += get_max_sub_group_size()) {
+                sumf += (float) x[i] * (float) y[i];
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        global float4 * x4 = (global float4 *)x;
+        for (int row = 0; row < N_F32_F32; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            ulong offset_src1 = r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+            global float  * y  = (global float  *) (src1 + offset_src1);
+            global float4 * y4 = (global float4 *) y;
+
+            float sumf = 0;
+            for (int i = get_sub_group_local_id(); i < ne00/4; i += get_max_sub_group_size()) {
+                sumf += (float) x4[i].s0 * y4[i].s0;
+                sumf += (float) x4[i].s1 * y4[i].s1;
+                sumf += (float) x4[i].s2 * y4[i].s2;
+                sumf += (float) x4[i].s3 * y4[i].s3;
+            }
+
+            float all_sum = sub_group_reduce_add(sumf);
+            if (get_sub_group_local_id() == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) {
+                    all_sum += (float) x[i] * y[i];
+                }
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
new file mode 100644
index 0000000..d50bd1f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl
@@ -0,0 +1,189 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_mv_mxfp4_f32(
+    global char * src0,
+    global char * src1,
+    global char * dst,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3,
+    local  char * shmem
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11 = idx % ne11;
+    int i12 = iid1;
+
+    int i1 = idx;
+    int i2 = i12;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11*nb11 + i12*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur,
+        ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl
new file mode 100644
index 0000000..f65e86e
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32_flat.cl
@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+
+static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) {
+    ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00;
+    fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00;
+    fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00;
+    fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800;
+    bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800;
+    bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800;
+    bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800;
+
+    fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo;
+    fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi;
+    fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo;
+    fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi;
+
+    sign_a.lo = (fp4x4 << 12) & 0x8000;
+    sign_a.hi = (fp4x4 << 8) & 0x8000;
+    sign_b.lo = (fp4x4 << 4) & 0x8000;
+    sign_b.hi = fp4x4 & 0x8000;
+
+    fp16_packed_a = sign_a + bias_a + fp16_packed_a;
+    fp16_packed_b = sign_b + bias_b + fp16_packed_b;
+
+    return as_half4((ushort4)(fp16_packed_a, fp16_packed_b));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 4
+#define N_SG_MXFP4 1
+#define N_SIMDWIDTH 64
+#define SRC0Q_IMG
+#endif
+
+kernel void kernel_mul_mv_id_mxfp4_f32_flat(
+#ifdef SRC0Q_IMG
+    __read_only image1d_buffer_t src0_q,
+#else
+    global uchar * src0_q,
+#endif
+    global uchar * src0_e,
+    global uchar * src1,
+    ulong         offset1,
+    global uchar * src2,
+    ulong         offset2,
+    global uchar * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    dst  = dst  + offsetd;
+
+    const int iid1 = get_group_id(2) / ne20;
+    const int idx  = get_group_id(2) % ne20;
+
+    uint i02 = ((global uint *) (src2 + offset2 + iid1 * nb21))[idx];
+
+    int i11 = idx % ne11;
+
+    int nb = ne00 / QK_MXFP4;
+
+    uint src0_off = i02*nb02;
+    src0_off /= 17; // 17 = sizeof(block_mxfp4)
+
+    src0_e = src0_e + src0_off;
+
+    dst = dst + (idx * ne0 + iid1 * ne1 * ne0) * sizeof(float);
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint offset_src0 = first_row*nb01;
+    offset_src0 /= 17; // 17 = sizeof(block_mxfp4)
+#ifdef SRC0Q_IMG
+    ulong offset_q = src0_off + offset_src0;
+#else
+    src0_q = src0_q + src0_off*16;
+    global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
+#endif
+    global uchar * x_e = src0_e + offset_src0;
+
+    const short ix = get_sub_group_local_id() >> 1;
+    const short it = get_sub_group_local_id() & 1;
+
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    src1 = src1 + offset1 + i11 * nb11 + iid1 * nb12;
+    global float * y   = (global float *) (src1 + r1 * nb11);
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH / 2) {
+        global float4 * y4 = (global float4 *)yb;
+
+        #pragma unroll
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            uchar xb_e = x_e[row * nb + ib];
+#ifdef SRC0Q_IMG
+            ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy);
+#else
+            ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it));
+#endif
+
+            half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0);
+            half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1);
+            float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2);
+            fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3);
+            acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH / 2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *)dst + (ulong)r1 * ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl
new file mode 100644
index 0000000..7ccf41e
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q4_0_f32_8x_flat.cl
@@ -0,0 +1,283 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global char  * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q4_0_f32_8x_flat(
+        global char  *  src0_q,
+        global half  *  src0_d,
+        global float *  src1,
+        ulong           offset1,
+        global char  *  src2,
+        ulong           offset2,
+        global float *  dst,
+        ulong           offsetd,
+        int             ne00,
+        int             ne01,
+        int             ne02,
+        ulong           nb00,
+        ulong           nb02,
+        int             ne10,
+        int             ne11,
+        int             ne12,
+        ulong           nb11,
+        ulong           nb12,
+        int             ne20,
+        int             ne21,
+        ulong           nb21,
+        int             ne0,
+        int             ne1,
+        int             r2,
+        int             r3
+) {
+    src1 = (global float *)((global char *)src1 + offset1);
+    src2 = (global char  *)((global char *)src2 + offset2);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    const int i02 = ((global int *)(src2 + iid1*nb21))[idx];
+
+    const int i11 = idx%ne11;
+    const int i12 = iid1;
+
+    const int i1 = idx;
+    const int i2 = i12;
+
+    global char  * src0_q_cur = src0_q + (i02*nb02/nb00)*(QK4_0/2);
+    global half  * src0_d_cur = src0_d + (i02*nb02/nb00);
+    global float * src1_cur   = (global float *)((global char *) src1  + i11*nb11 + i12*nb12);
+    global float * dst_cur    = dst + i1*ne0 + i2*ne1*ne0;
+
+    mul_vec_q_n_f32_8x_flat(src0_q_cur, src0_d_cur, src1_cur, dst_cur, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl
new file mode 100644
index 0000000..f37e83e
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32.cl
@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q8_0_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    int iid1 = get_group_id(2)/ne20;
+    int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11_ = idx % ne11;
+    int i12_ = iid1;
+
+    int i1 = idx;
+    int i2 = i12_;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11_*nb11 + i12_*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    ulong offset_src1 = r1*nb11;
+    global float * y  = (global float *) (src1_cur + offset_src1);
+
+    // pointers to src0 rows
+    global block_q8_0 * ax[N_R0_Q8_0];
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        ulong offset_src0 = (first_row + row)*nb01;
+        ax[row] = (global block_q8_0 *) ((global char *) src0_cur + offset_src0);
+    }
+
+    float yl[NB_Q8_0];
+    float sumf[N_R0_Q8_0] = { 0.f };
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (short row = 0; row < N_R0_Q8_0; row++) {
+            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
+            float sumq = 0.f;
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*ax[row][ib].d;
+        }
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        float tot = sub_group_reduce_add(sumf[row]);
+
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl
new file mode 100644
index 0000000..fd3a071
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_id_q8_0_f32_flat.cl
@@ -0,0 +1,222 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_q8_0_f32_flat(
+    global char * src0_q,
+    global half * src0_d,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1
+) {
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    int iid1 = (int)get_group_id(2)/ne20;
+    int idx  = (int)get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11_ = idx % ne11;
+    int i12_ = iid1;
+
+    int i1 = idx;
+    int i2 = i12_;
+
+    // 34 == sizeof(block_q8_0)
+    uint src0_off = i02*nb02;
+    src0_off /= 34;
+
+    global char * src0_q_cur = src0_q + src0_off*sizeof(char)*QK8_0;
+    global half * src0_d_cur = src0_d + src0_off;
+    global char * src1_cur   = src1 + i11_*nb11 + i12_*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    ulong offset_src1 = r1*nb11;
+    global float * y  = (global float *) (src1_cur + offset_src1);
+
+    // pointers to src0 rows
+    uint offset_src0_base = first_row*nb01;
+
+    global char * ax0, * ax1, * ax2, * ax3;
+    global half * ad0, * ad1, * ad2, * ad3;
+    uint offset_src0;
+
+    offset_src0 = offset_src0_base + 0*nb01;
+    offset_src0 = offset_src0/34;
+    ax0 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad0 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 1*nb01;
+    offset_src0 = offset_src0/34;
+    ax1 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad1 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 2*nb01;
+    offset_src0 = offset_src0/34;
+    ax2 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad2 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 3*nb01;
+    offset_src0 = offset_src0/34;
+    ax3 = (global char *) ((global char *) src0_q_cur + offset_src0*sizeof(char)*QK8_0);
+    ad3 = (global half *) ((global char *) src0_d_cur + offset_src0*sizeof(half));
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    float8 yl;
+    float8 qv;
+    float4 sumf = 0.f;
+    float  sumq = 0.f;
+    global char * qs;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        yl = vload8(0, yb);
+
+        qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s0 += sumq*ad0[ib];
+
+        qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s1 += sumq*ad1[ib];
+
+        qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s2 += sumq*ad2[ib];
+
+        qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s3 += sumq*ad3[ib];
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst_cur + (ulong)r1*ne0;
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst_f32[first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst_f32[first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst_f32[first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst_f32[first_row + 3] = tot.s3;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
new file mode 100644
index 0000000..9a4d4b9
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl
@@ -0,0 +1,144 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl
new file mode 100644
index 0000000..3d5a923
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32_flat.cl
@@ -0,0 +1,167 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+
+static inline half4 mxfp4_to_fp16_packed(ushort fp4x4) {
+    ushort2 fp16_packed_a, fp16_packed_b, bias_a, bias_b, sign_a, sign_b;
+    fp16_packed_a.lo = (fp4x4 << 9) & 0x0E00;
+    fp16_packed_a.hi = (fp4x4 << 5) & 0x0E00;
+    fp16_packed_b.lo = (fp4x4 << 1) & 0x0E00;
+    fp16_packed_b.hi = (fp4x4 >> 3) & 0x0E00;
+
+    bias_a.lo = (fp16_packed_a.lo == 0) ? 0x0 : 0x3800;
+    bias_a.hi = (fp16_packed_a.hi == 0) ? 0x0 : 0x3800;
+    bias_b.lo = (fp16_packed_b.lo == 0) ? 0x0 : 0x3800;
+    bias_b.hi = (fp16_packed_b.hi == 0) ? 0x0 : 0x3800;
+
+    fp16_packed_a.lo = (fp16_packed_a.lo == 0x0200) ? 0x0 : fp16_packed_a.lo;
+    fp16_packed_a.hi = (fp16_packed_a.hi == 0x0200) ? 0x0 : fp16_packed_a.hi;
+    fp16_packed_b.lo = (fp16_packed_b.lo == 0x0200) ? 0x0 : fp16_packed_b.lo;
+    fp16_packed_b.hi = (fp16_packed_b.hi == 0x0200) ? 0x0 : fp16_packed_b.hi;
+
+    sign_a.lo = (fp4x4 << 12) & 0x8000;
+    sign_a.hi = (fp4x4 << 8) & 0x8000;
+    sign_b.lo = (fp4x4 << 4) & 0x8000;
+    sign_b.hi = fp4x4 & 0x8000;
+
+    fp16_packed_a = sign_a + bias_a + fp16_packed_a;
+    fp16_packed_b = sign_b + bias_b + fp16_packed_b;
+
+    return as_half4((ushort4)(fp16_packed_a, fp16_packed_b));
+}
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+    bits = (x == 0) ? 0x00400000 : ((uint) x << 23);
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#define SRC0Q_IMG
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32_flat(
+#ifdef SRC0Q_IMG
+    __read_only image1d_buffer_t src0_q,
+#else
+    global uchar * src0_q,
+#endif
+    global uchar * src0_e,
+    global uchar * src1,
+    ulong          offset1,
+    global uchar * dst,
+    ulong          offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3
+) {
+    src1 = src1 + offset1;
+    dst = dst + offsetd;
+
+    int nb = ne00 / QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im % ne12;
+    uint i13 = im / ne12;
+
+    uint offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    // 17 = sizeof(block_mxfp4)
+    offset_src0 /= 17;
+#ifdef SRC0Q_IMG
+    ulong offset_q = offset_src0;
+#else
+    global uchar16 * x_q = (global uchar16 *)(src0_q) + offset_src0;
+#endif
+    global uchar * x_e = src0_e + offset_src0;
+
+    ulong offset_src1 = r1 * nb11 + i12 * nb12 + i13 * nb13;
+    global float * y = (global float *)(src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id() >> 1;  // 0...15
+    const short it = get_sub_group_local_id() & 1;  // 0 or 1
+
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+
+        #pragma unroll
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            uchar xb_e = x_e[row * nb + ib];
+#ifdef SRC0Q_IMG
+            ushort4 xb_q = as_ushort4(read_imageui(src0_q, (offset_q + row * nb + ib) * 2 + it).xy);
+#else
+            ushort4 xb_q = vload4(0, (global ushort *)((global uchar *)(x_q + row * nb + ib) + 8 * it));
+#endif
+
+            half4 fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s0);
+            half4 fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s1);
+            float4 acc1 = y4[0] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[4] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            fp16x4_0 = mxfp4_to_fp16_packed(xb_q.s2);
+            fp16x4_1 = mxfp4_to_fp16_packed(xb_q.s3);
+            acc1 += y4[1] * (float4)(fp16x4_0.s0, fp16x4_0.s2, fp16x4_1.s0, fp16x4_1.s2);
+            acc1 += y4[5] * (float4)(fp16x4_0.s1, fp16x4_0.s3, fp16x4_1.s1, fp16x4_1.s3);
+
+            sumf[row] += e8m0_to_fp32(xb_e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl
new file mode 100644
index 0000000..52141e0
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32.cl
@@ -0,0 +1,192 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//------------------------------------------------------------------------------
+// mul_vec_q_n_f32
+//------------------------------------------------------------------------------
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_4_0_dot_y(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        private float * yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float2 acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+    for (int i = 0; i < 8; i+=2) {
+        acc.s0 += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc.s1 += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (sumy * -8.f + acc.s0 + acc.s1);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];       // src1 vector cache
+    float sumf[N_DST]={0.f};
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
+        }
+
+        for (int row = 0; row < N_DST; row++) {
+            sumf[row] += block_q_4_0_dot_y(x+ib+row*nb, sumy, yl, il);
+        }
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float tot[N_DST] = {
+        sub_group_reduce_add(sumf[0]), sub_group_reduce_add(sumf[1]),
+        sub_group_reduce_add(sumf[2]), sub_group_reduce_add(sumf[3])};
+    for (int row = 0; row < N_DST; ++row) {
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot[row];
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl
new file mode 100644
index 0000000..3eebab8
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_16x_flat.cl
@@ -0,0 +1,307 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 16 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 16
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 16x output.
+// Eeach simdgroup outputs 16 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float16 sumf = (float16)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f,
+                             0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  0*nb*QK4_0/2, d + ib +  0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  1*nb*QK4_0/2, d + ib +  1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  2*nb*QK4_0/2, d + ib +  2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  3*nb*QK4_0/2, d + ib +  3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  4*nb*QK4_0/2, d + ib +  4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  5*nb*QK4_0/2, d + ib +  5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  6*nb*QK4_0/2, d + ib +  6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  7*nb*QK4_0/2, d + ib +  7*nb, sumy, yl, il);
+
+        sumf.s8 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  8*nb*QK4_0/2, d + ib +  8*nb, sumy, yl, il);
+        sumf.s9 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 +  9*nb*QK4_0/2, d + ib +  9*nb, sumy, yl, il);
+        sumf.sa += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 10*nb*QK4_0/2, d + ib + 10*nb, sumy, yl, il);
+        sumf.sb += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 11*nb*QK4_0/2, d + ib + 11*nb, sumy, yl, il);
+
+        sumf.sc += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 12*nb*QK4_0/2, d + ib + 12*nb, sumy, yl, il);
+        sumf.sd += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 13*nb*QK4_0/2, d + ib + 13*nb, sumy, yl, il);
+        sumf.se += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 14*nb*QK4_0/2, d + ib + 14*nb, sumy, yl, il);
+        sumf.sf += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 15*nb*QK4_0/2, d + ib + 15*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float16 tot = (float16)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7),
+
+        sub_group_reduce_add(sumf.s8), sub_group_reduce_add(sumf.s9),
+        sub_group_reduce_add(sumf.sa), sub_group_reduce_add(sumf.sb),
+        sub_group_reduce_add(sumf.sc), sub_group_reduce_add(sumf.sd),
+        sub_group_reduce_add(sumf.se), sub_group_reduce_add(sumf.sf)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+
+        if (first_row + 8 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 8] = tot.s8;
+        }
+        if (first_row + 9 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 9] = tot.s9;
+        }
+        if (first_row + 10 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 10] = tot.sa;
+        }
+        if (first_row + 11 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 11] = tot.sb;
+        }
+
+        if (first_row + 12 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 12] = tot.sc;
+        }
+        if (first_row + 13 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 13] = tot.sd;
+        }
+        if (first_row + 14 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 14] = tot.se;
+        }
+        if (first_row + 15 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 15] = tot.sf;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_16x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_16x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl
new file mode 100644
index 0000000..38024d0
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_1d_8x_flat.cl
@@ -0,0 +1,265 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+inline float mm_block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows (in weights matrix)
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+//
+// This variant performs 1d blocking with 8x output.
+// Eeach simdgroup outputs 8 values on `n0` dim (row in the output matrix).
+//
+inline void mul_mat_q_n_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const int nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = (float8)(0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += mm_block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_1d_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_mat_q_n_f32_1d_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl
new file mode 100644
index 0000000..aed1ce7
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_8x_flat.cl
@@ -0,0 +1,272 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+// This function requires the original shuffled weights.
+// As a reminder, the original weights are shuffled so that (q[0], q[16]) are
+// packed together in a byte, so are (q[1], q[17]) and so on.
+inline float block_q_4_0_dot_y_flat(
+        global uchar * x,
+        global half  * dh,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float           d   = *dh;
+    global ushort * qs  = ((global ushort *)x + il/2);
+    float           acc = 0.f;
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+//
+// This variant outputs 8 values.
+//
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 8 // each SIMD group works on 8 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 32
+#elif defined (ADRENO_GPU)
+#define N_DST 8
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is the linear global id of
+    // a SIMD group in the grid. Each SIMD group produces N_DST values in the
+    // result, hence uses nb blocks, i.e., the offset becomes first_row*nb.
+    // Currently with llama2 7B, im is always 0.
+    // TODO: how to handle im/gqa*(nb*ne0)?
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    // The number of scales is the same as the number of blocks.
+    ulong offset0_d = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    // Each block contains QK4_0/2 uchars, hence offset for qs is as follows.
+    ulong offset0_q = (first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02)) * QK4_0/2;
+
+    global uchar * x = (global uchar *) src0_q + offset0_q;
+    global half  * d = (global half  *) src0_d + offset0_d;
+    global float * y = (global float *) src1   + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;
+    float8 sumf = 0.f;
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix*QK4_0 + il;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0.f;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 0*nb*QK4_0/2, d + ib + 0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 1*nb*QK4_0/2, d + ib + 1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 2*nb*QK4_0/2, d + ib + 2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 3*nb*QK4_0/2, d + ib + 3*nb, sumy, yl, il);
+
+        sumf.s4 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 4*nb*QK4_0/2, d + ib + 4*nb, sumy, yl, il);
+        sumf.s5 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 5*nb*QK4_0/2, d + ib + 5*nb, sumy, yl, il);
+        sumf.s6 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 6*nb*QK4_0/2, d + ib + 6*nb, sumy, yl, il);
+        sumf.s7 += block_q_4_0_dot_y_flat(x + ib*QK4_0/2 + 7*nb*QK4_0/2, d + ib + 7*nb, sumy, yl, il);
+
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    float8 tot = (float8)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3),
+        sub_group_reduce_add(sumf.s4), sub_group_reduce_add(sumf.s5),
+        sub_group_reduce_add(sumf.s6), sub_group_reduce_add(sumf.s7)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+
+        if (first_row + 4 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 4] = tot.s4;
+        }
+        if (first_row + 5 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 5] = tot.s5;
+        }
+        if (first_row + 6 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 6] = tot.s6;
+        }
+        if (first_row + 7 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 7] = tot.s7;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_8x_flat(
+        global uchar * src0_q,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_8x_flat(src0_q, src0_d, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl
new file mode 100644
index 0000000..9295521
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_0_f32_v.cl
@@ -0,0 +1,254 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q4_0
+//------------------------------------------------------------------------------
+struct block_q4_0
+{
+    half d;
+    uint8_t qs[QK4_0 / 2];
+};
+
+//
+// This variant unrolls the loops and uses vector types instead of pointers.
+// It improves performance on Adreno but not so much on Intel.
+//
+inline float block_q_4_0_dot_y_v(
+        global struct block_q4_0 * qb_curr,
+        float sumy,
+        float16 yl,
+        int il
+) {
+    float d = qb_curr->d;
+    float acc = 0.f;
+    global ushort * qs = ((global ushort *)qb_curr + 1 + il/2);
+
+    acc += yl.s0 * (qs[0] & 0x000F);
+    acc += yl.s1 * (qs[0] & 0x0F00);
+    acc += yl.s8 * (qs[0] & 0x00F0);
+    acc += yl.s9 * (qs[0] & 0xF000);
+
+    acc += yl.s2 * (qs[1] & 0x000F);
+    acc += yl.s3 * (qs[1] & 0x0F00);
+    acc += yl.sa * (qs[1] & 0x00F0);
+    acc += yl.sb * (qs[1] & 0xF000);
+
+    acc += yl.s4 * (qs[2] & 0x000F);
+    acc += yl.s5 * (qs[2] & 0x0F00);
+    acc += yl.sc * (qs[2] & 0x00F0);
+    acc += yl.sd * (qs[2] & 0xF000);
+
+    acc += yl.s6 * (qs[3] & 0x000F);
+    acc += yl.s7 * (qs[3] & 0x0F00);
+    acc += yl.se * (qs[3] & 0x00F0);
+    acc += yl.sf * (qs[3] & 0xF000);
+
+    return d * (sumy * -8.f + acc);
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // assuming SIMD group size is 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_vec_q_n_f32_v(
+        global void * src0,
+        global float * src1,
+        global float * dst,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    const ulong nb = ne00/QK4_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    // (r0 * N_SIMDGROUP + get_sub_group_id()) is essenatially the linear global
+    // id of a SIMD group in the grid.
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global struct block_q4_0 * x = (global struct block_q4_0 *) src0 + offset0;
+    global float             * y = (global float             *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float16 yl;       // src1 vector cache
+    float4 sumf = (float4)(0.f, 0.f, 0.f, 0.f);
+
+    int ix = get_sub_group_local_id()/2;
+    int il = 8*(get_sub_group_local_id()%2);
+
+    global float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        float sumy = 0;
+
+        sumy += yb[0];
+        sumy += yb[1];
+        sumy += yb[2];
+        sumy += yb[3];
+        sumy += yb[4];
+        sumy += yb[5];
+        sumy += yb[6];
+        sumy += yb[7];
+
+        sumy += yb[16];
+        sumy += yb[17];
+        sumy += yb[18];
+        sumy += yb[19];
+        sumy += yb[20];
+        sumy += yb[21];
+        sumy += yb[22];
+        sumy += yb[23];
+
+
+        yl.s0 = yb[0];
+        yl.s1 = yb[1]/256.f;
+
+        yl.s2 = yb[2];
+        yl.s3 = yb[3]/256.f;
+
+        yl.s4 = yb[4];
+        yl.s5 = yb[5]/256.f;
+
+        yl.s6 = yb[6];
+        yl.s7 = yb[7]/256.f;
+
+        yl.s8 = yb[16]/16.f;
+        yl.s9 = yb[17]/4096.f;
+
+        yl.sa = yb[18]/16.f;
+        yl.sb = yb[19]/4096.f;
+
+        yl.sc = yb[20]/16.f;
+        yl.sd = yb[21]/4096.f;
+
+        yl.se = yb[22]/16.f;
+        yl.sf = yb[23]/4096.f;
+
+        sumf.s0 += block_q_4_0_dot_y_v(x+ib+0*nb, sumy, yl, il);
+        sumf.s1 += block_q_4_0_dot_y_v(x+ib+1*nb, sumy, yl, il);
+        sumf.s2 += block_q_4_0_dot_y_v(x+ib+2*nb, sumy, yl, il);
+        sumf.s3 += block_q_4_0_dot_y_v(x+ib+3*nb, sumy, yl, il);
+
+        // One thread in a SIMD group (i.e., subgroup) handles a half block,
+        // hence then entire SIMD group handles SIMDWIDTH/2 blocks.
+        // y points to the activation matrix (of type float). Therefore for
+        // one thread, the # of blocks y should advance is SIMDWIDTH/2 (because
+        // SIMDWIDTH/2 blocks are processed by a SIMD group) - in terms of
+        // floats, it is QK4_0 * (SIMDWIDTH/2), where QK4_0 is the block size.
+        yb += QK4_0 * (N_SIMDWIDTH/2);
+    }
+
+    // The above does not work for Adreno - it produces incorrect results for
+    // row = 1, 2, 3 and only row = 0 gives the correct result.
+    // If N_DST is changed, the below array must be initialized accordingly.
+    // This also seems to perform better on Intel.
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0), sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2), sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_q4_0_f32_v(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    mul_vec_q_n_f32_v(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl
new file mode 100644
index 0000000..71ab989
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q4_k_f32.cl
@@ -0,0 +1,180 @@
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// block_q4_K
+//------------------------------------------------------------------------------
+#define QK_K            256
+#define K_SCALE_SIZE    12
+
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+typedef struct {
+    half d;    // super-block scale for quantized scales
+    half dmin; // super-block scale for quantized mins
+
+    uchar scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uchar qs[QK_K/2];           // 4-bit quants
+} block_q4_K;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// number of (super) blocks each subgroup processes
+// each thread in a subgroup processes a block (32 weights)
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q4_K_f32(
+        global char * src0,
+        int offset0,
+        global char * src1,
+        int offset1,
+        global char * dst,
+        int offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    ushort kmask1 = 0x3f3f;
+    ushort kmask2 = 0x0f0f;
+    ushort kmask3 = 0xc0c0;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // block index (inside super block)
+    int iq = it/4;     // 0 or 1 - first or second half of the super block
+    int ir = it%4;     // 0...3 - block index in the half super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    int offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_q4_K * x = (global block_q4_K *) (src0 + offset_src0);
+    global float      * y = (global float      *) (src1 + offset_src1);
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    ushort  sc16[4];
+    uchar * sc8 = (uchar *)sc16;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+0];
+            sumy.s0 += yl[i+0];
+
+            yl[i+8] = y4[i+32];
+            sumy.s1 += yl[i+8];
+
+            yh[i+0] = y4[i+128];
+            sumy.s2 += yh[i+0];
+
+            yh[i+8] = y4[i+160];
+            sumy.s3 += yh[i+8];
+        }
+
+        global ushort * sc = (global ushort *)x[ib].scales + iq;
+        global ushort * q1 = (global ushort *)x[ib].qs + 16 * iq + 4 * ir;
+        global half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            global ushort * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1.s0 += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1.s1 += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1.s2 += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1.s3 += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2.s0 += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2.s1 += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2.s2 += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2.s3 += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1.s0 + 1.f/256.f * acc1.s1) * sc8[0] +
+                                 (acc1.s2 + 1.f/256.f * acc1.s3) * sc8[1] * 1.f/16.f +
+                                 (acc2.s0 + 1.f/256.f * acc2.s1) * sc8[4] +
+                                 (acc2.s2 + 1.f/256.f * acc2.s3) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy.s0 * sc8[2] + sumy.s1 * sc8[3] + sumy.s2 * sc8[6] + sumy.s3 * sc8[7]);
+
+            q1 += nb01/2;
+            sc += nb01/2;
+            dh += nb01/2;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
new file mode 100644
index 0000000..819e519
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32.cl
@@ -0,0 +1,194 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK4_0                   32
+#define QR4_0                   2
+#define QK4_1                   32
+#define QR4_1                   2
+#define QK5_0                   32
+#define QR5_0                   2
+#define QK5_1                   32
+#define QR5_1                   2
+#define QK8_0                   32
+#define QR8_0                   1
+#define QK_K                    256
+#define K_QUANTS_PER_ITERATION  2
+
+typedef char int8_t;
+typedef uchar uint8_t;
+typedef short int16_t;
+typedef ushort uint16_t;
+typedef int int32_t;
+typedef uint uint32_t;
+
+//------------------------------------------------------------------------------
+// block_q6_K
+//------------------------------------------------------------------------------
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    half d;             // super-block scale
+} block_q6_K;
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32
+//------------------------------------------------------------------------------
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 1 // number of rows each SIMD group works on
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 1
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32(
+        global void * src0,
+        ulong offset0,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    uchar kmask1 = 0x03;
+    uchar kmask2 = 0x0C;
+    uchar kmask3 = 0x30;
+    uchar kmask4 = 0xC0;
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int row = N_SIMDGROUP * r0 + get_sub_group_id();
+
+    if (row >= ne01) {
+        return;
+    }
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    global block_q6_K * x = (global block_q6_K *) src0 + row*nb + offset_src0;
+    global float      * yy = (global float     *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float sumf = 0;
+
+    // For Q6_K quantization, 16 values forms a subblock, 16 subblock forms a
+    // block. Values in a subblock shares a scale that is quantized with 8 bits;
+    // the entire block shares a single floating point scale.
+    // For work distribution, each thread processes a subblock (16 weights), hence
+    // 16 threads process a (super) block -- a subgroup thus handles SIMDWIDTH/16
+    // (super) blocks -- this is the block stride.
+    // The 16 threads that process a (super) block are split into 2 portions, each has
+    // 8 threads; each portion works on 8 subblocks.
+    // For subgroup of 16 threads, the entire subgroup works on a single (super) block
+    // before moving to the next (super) block. Thread0 - thread7 work on the
+    // first 8 subblocks; thread8 - thread15 works on the last 8 subblocks.
+    // Thread0 - thread3 work on subblocks 0, 2, 4, 6; thread4 - thread7 work on
+    // subblocks 1, 3, 5, 7. Each thread does not work on an entire subblock, but
+    // works on a total of 16 weight values.
+    int tid  = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix   = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip   = tid/8;   // first or second half of (super) block (0 or 1)
+    int il   = tid%8;   // each half has 8 parts, one per scale
+    int n    = 4;       // 4 scales at a time (and 4 sums)
+    int l0   = n*il;    // offset into half-block, 0..28
+    int is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    int y_offset = 128*ip + l0;
+    int q_offset_l = 64*ip + l0;
+    int q_offset_h = 32*ip + l0;
+
+    for (int i = ix; i < nb; i += BLOCK_STRIDE) {
+
+        global uint8_t * q1 = x[i].ql + q_offset_l;
+        global uint8_t * q2 = q1 + QK_K/8;
+        global uint8_t * qh = x[i].qh + q_offset_h;
+        global int8_t  * sc = x[i].scales + is;
+
+        global float * y = yy + i * QK_K + y_offset;
+
+        float dall = x[i].d;
+
+        float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+        sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & kmask4) >> 2)) - 32.f);
+
+        sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & kmask1) << 4)) - 32.f);
+        sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & kmask2) << 2)) - 32.f);
+        sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & kmask3) << 0)) - 32.f);
+        sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & kmask4) >> 2)) - 32.f);
+
+        sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+    }
+
+    float tot = sub_group_reduce_add(sumf);
+    if (get_sub_group_local_id() == 0) {
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl
new file mode 100644
index 0000000..86fe09c
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q6_k_f32_flat.cl
@@ -0,0 +1,194 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// kernel_mul_mv_q6_K_f32_flat
+//------------------------------------------------------------------------------
+#define Q6_K_MASK1 0x03
+#define Q6_K_MASK2 0x0C
+#define Q6_K_MASK3 0x30
+#define Q6_K_MASK4 0xC0
+
+#define QK_K       256
+
+inline float block_q_6_K_dot_y_flat(
+    global uchar * blk_ql,
+    global uchar * blk_qh,
+    global char  * blk_scales,
+    global half  * blk_d,
+    global float * yy,
+    int ib,
+    int ip,
+    int is,
+    int l0
+) {
+    int y_offset   = 128*ip + l0;
+    int q_offset_l =  64*ip + l0;
+    int q_offset_h =  32*ip + l0;
+
+    global uchar * q1 = blk_ql     + ib*128 + q_offset_l;
+    global uchar * q2 = q1         + QK_K/8;
+    global uchar * qh = blk_qh     + ib*64 + q_offset_h;
+    global char  * sc = blk_scales + ib*16 + is;
+
+    global float * y = yy + ib * QK_K + y_offset;
+
+    float dall = blk_d[ib];
+
+    float  sumf = 0;
+    float4 sums = {0.f, 0.f, 0.f, 0.f};
+
+    sums.s0 += y[0+ 0] * ((float)((q1[0] & 0xF) | ((qh[0] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[0+32] * ((float)((q2[0] & 0xF) | ((qh[0] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[0+64] * ((float)((q1[0]  >> 4) | ((qh[0] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[0+96] * ((float)((q2[0]  >> 4) | ((qh[0] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[1+ 0] * ((float)((q1[1] & 0xF) | ((qh[1] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[1+32] * ((float)((q2[1] & 0xF) | ((qh[1] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[1+64] * ((float)((q1[1]  >> 4) | ((qh[1] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[1+96] * ((float)((q2[1]  >> 4) | ((qh[1] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[2+ 0] * ((float)((q1[2] & 0xF) | ((qh[2] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[2+32] * ((float)((q2[2] & 0xF) | ((qh[2] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[2+64] * ((float)((q1[2]  >> 4) | ((qh[2] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[2+96] * ((float)((q2[2]  >> 4) | ((qh[2] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sums.s0 += y[3+ 0] * ((float)((q1[3] & 0xF) | ((qh[3] & Q6_K_MASK1) << 4)) - 32.f);
+    sums.s1 += y[3+32] * ((float)((q2[3] & 0xF) | ((qh[3] & Q6_K_MASK2) << 2)) - 32.f);
+    sums.s2 += y[3+64] * ((float)((q1[3]  >> 4) | ((qh[3] & Q6_K_MASK3) << 0)) - 32.f);
+    sums.s3 += y[3+96] * ((float)((q2[3]  >> 4) | ((qh[3] & Q6_K_MASK4) >> 2)) - 32.f);
+
+    sumf += dall * (sums.s0 * sc[0] + sums.s1 * sc[2] + sums.s2 * sc[4] + sums.s3 * sc[6]);
+
+    return sumf;
+}
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 2
+#define N_SIMDWIDTH 64
+#endif
+
+#define BLOCK_STRIDE (N_SIMDWIDTH/16) // number of blocks each subgroup processes
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q6_K_f32_flat(
+        global uchar * src0_ql,
+        global uchar * src0_qh,
+        global char  * src0_s,
+        global half  * src0_d,
+        global float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne10,
+        int ne12,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    int first_row = (N_SIMDGROUP * r0 + get_sub_group_id()) * N_DST;
+
+    ulong offset_src0    = first_row*nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    ulong offset_src0_ql = offset_src0 * 128;
+    ulong offset_src0_qh = offset_src0 * 64;
+    ulong offset_src0_s  = offset_src0 * 16;
+    ulong offset_src0_d  = offset_src0;
+
+    global uchar * blk_ql     = (global uchar *) src0_ql + offset_src0_ql;
+    global uchar * blk_qh     = (global uchar *) src0_qh + offset_src0_qh;
+    global char  * blk_scales = (global char  *) src0_s  + offset_src0_s;
+    global half  * blk_d      = (global half  *) src0_d  + offset_src0_d;
+    global float * yy         = (global float *) src1    + r1*ne10 + im*ne00*ne1;
+
+    int tid = get_sub_group_local_id()/BLOCK_STRIDE; // first block_stride groups have tid=0
+    int ix  = get_sub_group_local_id()%BLOCK_STRIDE; // first block is 0..block_stride-1
+    int ip  = tid/8;   // first or second half of (super) block (0 or 1)
+    int il  = tid%8;   // each half has 8 parts, one per scale
+    int n   = 4;       // 4 scales at a time (and 4 sums)
+    int l0  = n*il;    // offset into half-block, 0..28
+    int is  = 8*ip + l0/16; // 0, 1, 8, 9
+
+    float4 sumf = 0;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        if (first_row + 0 < ne01) {
+            sumf.s0 += block_q_6_K_dot_y_flat(blk_ql + 0*nb*128, blk_qh + 0*nb*64, blk_scales + 0*nb*16, blk_d + 0*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 1 < ne01) {
+            sumf.s1 += block_q_6_K_dot_y_flat(blk_ql + 1*nb*128, blk_qh + 1*nb*64, blk_scales + 1*nb*16, blk_d + 1*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 2 < ne01) {
+            sumf.s2 += block_q_6_K_dot_y_flat(blk_ql + 2*nb*128, blk_qh + 2*nb*64, blk_scales + 2*nb*16, blk_d + 2*nb, yy, ib, ip, is, l0);
+        }
+        if (first_row + 3 < ne01) {
+            sumf.s3 += block_q_6_K_dot_y_flat(blk_ql + 3*nb*128, blk_qh + 3*nb*64, blk_scales + 3*nb*16, blk_d + 3*nb, yy, ib, ip, is, l0);
+        }
+    }
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + 3] = tot.s3;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl
new file mode 100644
index 0000000..7e88c74
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32.cl
@@ -0,0 +1,125 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q8_0_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13;
+    global float * y  = (global float *) (src1 + offset_src1);
+
+    // pointers to src0 rows
+    global block_q8_0 * ax[N_R0_Q8_0];
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        ulong offset_src0 = (first_row + row)*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+        ax[row] = (global block_q8_0 *) ((global char *) src0 + offset_src0);
+    }
+
+    float yl[NB_Q8_0];
+    float sumf[N_R0_Q8_0] = { 0.f };
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        for (short i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (short row = 0; row < N_R0_Q8_0; row++) {
+            global char * qs = ax[row][ib].qs + il*NB_Q8_0;
+            float sumq = 0.f;
+            for (short iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*ax[row][ib].d;
+        }
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_Q8_0; ++row) {
+        float tot = sub_group_reduce_add(sumf[row]);
+
+        if (get_sub_group_local_id() == 0 && first_row + row < ne01) {
+            dst_f32[first_row + row] = tot;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl
new file mode 100644
index 0000000..71d159f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/mul_mv_q8_0_f32_flat.cl
@@ -0,0 +1,202 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK8_0 32
+typedef struct {
+    half d;       // delta
+    char qs[QK8_0]; // quants
+} block_q8_0;
+
+#define NB_Q8_0 8
+
+#ifdef INTEL_GPU
+#define N_R0_Q8_0 4 // number of rows each subgroup works on
+#define N_SG_Q8_0 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_Q8_0 4
+#define N_SG_Q8_0 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q8_0_f32_flat(
+    global char * src0_q,
+    global half * src0_d,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    int           ne01,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3
+) {
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    int nb = ne00/QK8_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0*N_SG_Q8_0 + get_sub_group_id()) * N_R0_Q8_0;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src1 = r1*nb11 + i12*nb12 + i13*nb13;
+    global float * y  = (global float *) (src1 + offset_src1);
+
+    // pointers to src0 rows
+    uint offset_src0_base = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+
+    global char * ax0, * ax1, * ax2, * ax3;
+    global half * ad0, * ad1, * ad2, * ad3;
+    uint offset_src0;
+
+    offset_src0 = offset_src0_base + 0*nb01;
+    offset_src0 = offset_src0/34;
+    ax0 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad0 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 1*nb01;
+    offset_src0 = offset_src0/34;
+    ax1 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad1 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 2*nb01;
+    offset_src0 = offset_src0/34;
+    ax2 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad2 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    offset_src0 = offset_src0_base + 3*nb01;
+    offset_src0 = offset_src0/34;
+    ax3 = (global char *) ((global char *) src0_q + offset_src0*sizeof(char)*QK8_0);
+    ad3 = (global half *) ((global char *) src0_d + offset_src0*sizeof(half));
+
+    const short ix = get_sub_group_local_id()/4;
+    const short il = get_sub_group_local_id()%4;
+
+    global float * yb = y + ix*QK8_0 + il*NB_Q8_0;
+
+    float8 yl;
+    float8 qv;
+    float4 sumf = 0.f;
+    float  sumq = 0.f;
+    global char * qs;
+
+    // each thread handles NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/4) {
+        yl = vload8(0, yb);
+
+        qs = ax0 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s0 += sumq*ad0[ib];
+
+        qs = ax1 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s1 += sumq*ad1[ib];
+
+        qs = ax2 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s2 += sumq*ad2[ib];
+
+        qs = ax3 + ib*sizeof(char)*QK8_0 + il*NB_Q8_0;
+        qv = convert_float8(vload8(0, qs));
+        sumq = 0;
+        sumq += qv.s0*yl.s0;
+        sumq += qv.s1*yl.s1;
+        sumq += qv.s2*yl.s2;
+        sumq += qv.s3*yl.s3;
+        sumq += qv.s4*yl.s4;
+        sumq += qv.s5*yl.s5;
+        sumq += qv.s6*yl.s6;
+        sumq += qv.s7*yl.s7;
+        sumf.s3 += sumq*ad3[ib];
+
+        yb += N_SIMDWIDTH*NB_Q8_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    float4 tot = (float4)(
+        sub_group_reduce_add(sumf.s0),
+        sub_group_reduce_add(sumf.s1),
+        sub_group_reduce_add(sumf.s2),
+        sub_group_reduce_add(sumf.s3)
+    );
+
+    if (get_sub_group_local_id() == 0) {
+        if (first_row + 0 < ne01) {
+            dst_f32[first_row + 0] = tot.s0;
+        }
+        if (first_row + 1 < ne01) {
+            dst_f32[first_row + 1] = tot.s1;
+        }
+        if (first_row + 2 < ne01) {
+            dst_f32[first_row + 2] = tot.s2;
+        }
+        if (first_row + 3 < ne01) {
+            dst_f32[first_row + 3] = tot.s3;
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl
new file mode 100644
index 0000000..170f822
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/norm.cl
@@ -0,0 +1,161 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// norm
+//------------------------------------------------------------------------------
+kernel void kernel_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global void*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float * x = (global float *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+
+    // MEAN
+    // parallel sum
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        sum[get_local_id(0)] += x[i00];
+    }
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float mean  = sum[0] / ne00;
+
+    // recenter and VARIANCE
+    barrier(CLK_LOCAL_MEM_FENCE);
+    global float * y = dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+    sum[get_local_id(0)] = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = x[i00] - mean;
+        sum[get_local_id(0)] += y[i00] * y[i00];
+    }
+
+    // reduce
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint i = get_local_size(0)/2; i > 0; i /= 2) {
+        if (get_local_id(0) < i) {
+            sum[get_local_id(0)] += sum[get_local_id(0) + i];
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    float variance = sum[0] / ne00;
+
+    float scale = 1.0f/sqrt(variance + eps);
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        y[i00] = y[i00] * scale;
+    }
+}
+
+//------------------------------------------------------------------------------
+// norm_mul_add
+//------------------------------------------------------------------------------
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_norm_mul_add(
+        global char * src0_ptr, ulong src0_offset,
+        global char * src1_ptr, ulong src1_offset,
+        global char * src2_ptr, ulong src2_offset,
+        global char * dst_ptr,  ulong dst_offset,
+        int ne00, int ne01, int ne02, int ne03,
+        ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12, int ne13,
+        ulong nb11, ulong nb12, ulong nb13,
+        int ne20, int ne21, int ne22, int ne23,
+        ulong nb21, ulong nb22, ulong nb23,
+        ulong nbd1, ulong nbd2, ulong nbd3,
+        float eps,
+        local float2 * sums
+) {
+    const int i03 = get_group_id(2);
+    const int i02 = get_group_id(1);
+    const int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *)(src0_ptr + src0_offset + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * w = (global float4 *)(src1_ptr + src1_offset + (i01%ne11)*nb11 + (i02%ne12)*nb12 + (i03%ne13)*nb13);
+    global float4 * b = (global float4 *)(src2_ptr + src2_offset + (i01%ne21)*nb21 + (i02%ne22)*nb22 + (i03%ne23)*nb23);
+    global float4 * y = (global float4 *)(dst_ptr  + dst_offset  + i01*nbd1 + i02*nbd2 + i03*nbd3);
+
+    float p_sum = 0.0f;
+    float p_sum_sq = 0.0f;
+
+    const int n_chunks = ne00 / 4;
+    for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) {
+        float4 val = x[i00];
+        p_sum += val.x + val.y + val.z + val.w;
+        p_sum_sq += dot(val, val);
+    }
+
+    p_sum = sub_group_reduce_add(p_sum);
+    p_sum_sq = sub_group_reduce_add(p_sum_sq);
+
+    if (get_sub_group_local_id() == 0) {
+        sums[get_sub_group_id()] = (float2)(p_sum, p_sum_sq);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (get_local_id(0) == 0) {
+        float sum = 0.0f;
+        float sum_sq = 0.0f;
+        for (uint i = 0; i < get_num_sub_groups(); ++i) {
+            float2 s = sums[i];
+            sum += s.x;
+            sum_sq += s.y;
+        }
+
+        const float inv_ne00 = 1.0f / (float)ne00;
+        const float mean = sum * inv_ne00;
+        const float variance = mad(-mean, mean, sum_sq * inv_ne00);
+
+        sums[0] = (float2)(mean, rsqrt(variance + eps));
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float2 mean_scale = sums[0];
+    const float mean = mean_scale.x;
+    const float scale = mean_scale.y;
+    const float neg_mean_scale = -mean * scale;
+
+    for (int i00 = get_local_id(0); i00 < n_chunks; i00 += get_local_size(0)) {
+        const int w_idx = ne10 > 1 ? i00 : 0;
+        const int b_idx = ne20 > 1 ? i00 : 0;
+        const float4 norm_x = mad(x[i00], (float4)scale, (float4)neg_mean_scale);
+        y[i00] = mad(norm_x, w[w_idx], b[b_idx]);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl
new file mode 100644
index 0000000..31fb7cc
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/pad.cl
@@ -0,0 +1,39 @@
+kernel void kernel_pad(
+        global void * src0,
+        ulong offset0,
+        global void * dst,
+        ulong offsetd,
+        int ne00, int ne01, int ne02, int ne03,
+        ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+        int ne0, int ne1, int ne2, int ne3,
+        ulong nb0, ulong nb1, ulong nb2, ulong nb3,
+        int lp0, int rp0,
+        int lp1, int rp1,
+        int lp2, int rp2,
+        int lp3, int rp3
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst  + offsetd);
+
+    int i0 = get_global_id(0);
+    int i1 = get_group_id(1);
+    int i2 = get_group_id(2) % ne2;
+    int i3 = get_group_id(2) / ne2;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    uint src0_idx = (i3 - lp3)*nb03 + (i2 - lp2)*nb02 + (i1 - lp1)*nb01 + (i0 - lp0)*nb00;
+    uint dst_idx  =         i3*nb3  +         i2*nb2  +         i1*nb1  +         i0*nb0;
+
+    global float * src0_ptr = (global float *)((global char *)src0 + src0_idx);
+    global float * dst_ptr  = (global float *)((global char *)dst  + dst_idx);
+
+    bool in_src_bounds = (i0 >= lp0 && i0 < ne0 - rp0) &&
+                         (i1 >= lp1 && i1 < ne1 - rp1) &&
+                         (i2 >= lp2 && i2 < ne2 - rp2) &&
+                         (i3 >= lp3 && i3 < ne3 - rp3);
+
+    *dst_ptr = in_src_bounds ? *src0_ptr : 0.0f;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl
new file mode 100644
index 0000000..60ff28a
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/relu.cl
@@ -0,0 +1,16 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// relu
+//------------------------------------------------------------------------------
+kernel void kernel_relu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = fmax(0.0f, src0[get_global_id(0)]);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl
new file mode 100644
index 0000000..53951a5
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/repeat.cl
@@ -0,0 +1,38 @@
+kernel void kernel_repeat_f32(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int     ne00,
+        int     ne01,
+        int     ne02,
+        int     ne03,
+        ulong   nb00,
+        ulong   nb01,
+        ulong   nb02,
+        ulong   nb03,
+        int     ne0,
+        ulong   nb0,
+        ulong   nb1,
+        ulong   nb2,
+        ulong   nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst  + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    const int i03 = i3%ne03;
+    const int i02 = i2%ne02;
+    const int i01 = i1%ne01;
+
+    global const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i00 = i0%ne00;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i00*nb00));
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl
new file mode 100644
index 0000000..4b18d17
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/rms_norm.cl
@@ -0,0 +1,190 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+//------------------------------------------------------------------------------
+// rms_norm
+//------------------------------------------------------------------------------
+// This kernel depends on subgroup size.
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_rms_norm(
+        global void * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        float eps,
+        local float * sum // Note, the size depends on number of subgroups
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) ((global char *) src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float * x_scalar = (global float *) x;
+    float4 sumf = 0;
+    float all_sum = 0;
+
+    // parallel sum
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += x[i00] * x[i00];
+    }
+    all_sum = sumf.s0 + sumf.s1 + sumf.s2 + sumf.s3;
+    all_sum = sub_group_reduce_add(all_sum);
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = all_sum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+    // broadcast
+    for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
+       if (get_local_id(0) < i) {
+           sum[get_local_id(0)] += sum[get_local_id(0) + i];
+       }
+    }
+    if (get_local_id(0) == 0) {
+        for (int i = 4 * (ne00 / 4); i < ne00; i++) {
+            sum[0] += x_scalar[i];
+        }
+        sum[0] /= ne00;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    const float mean  = sum[0];
+    const float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    global float * y_scalar = (global float *) y;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = x[i00] * scale;
+    }
+    if (get_local_id(0) == 0) {
+        for (int i00 = 4 * (ne00 / 4); i00 < ne00; i00++) {
+            y_scalar[i00] = x_scalar[i00] * scale;
+        }
+    }
+}
+
+//------------------------------------------------------------------------------
+// rms_norm_mul
+//------------------------------------------------------------------------------
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_rms_norm_mul(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float eps,
+        local float * sum
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    // The size of sum is sizeof(float)*subgroup_size.
+    // Each subgroup writes its partial sum to this array.
+    // So the number of subgroups per workgroup for this kernel cannot exceed the subgroup size.
+    // This is generally true -
+    // for subgroup size 64, workgroup size should be less than 4096 (the max is usually 1024).
+    if (get_sub_group_id() == 0) {
+        sum[get_sub_group_local_id()] = 0.0f;
+    }
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float4 * f = (global float4 *) (src1 + (i03%ne13)*nb13 + (i02%ne12)*nb12 + (i01%ne11)*nb11);
+
+    float sumf = 0;
+
+    // parallel sum
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = sumf;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    //for (uint i = get_local_size(0) / get_max_sub_group_size() / 2; i > 0; i /= 2) {
+    //   if (get_local_id(0) < i) {
+    //       sum[get_local_id(0)] += sum[get_local_id(0) + i];
+    //   }
+    //}
+    //if (get_local_id(0) == 0) {
+    //    sum[0] /= ne00;
+    //}
+
+    //barrier(CLK_LOCAL_MEM_FENCE);
+
+    sumf = sum[get_sub_group_local_id()];
+    sumf = sub_group_reduce_add(sumf);
+
+    float mean  = sumf / ne00;
+    float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1);
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = (x[i00] * scale) * f[i00%(ne10/4)];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl
new file mode 100644
index 0000000..82f4cd8
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/rope.cl
@@ -0,0 +1,747 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// kernel_rope
+//------------------------------------------------------------------------------
+float rope_yarn_ramp(float low, float high, int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+float2 rope_yarn(
+    float theta_extrap, float freq_scale, float2 corr_dims, int i0, float ext_factor, float mscale
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.s0, corr_dims.s1, i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    return (float2)(cos(theta) * mscale, sin(theta) * mscale);
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+float2 rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow
+) {
+    // start and end correction dims
+    return (float2)(
+        max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base))),
+        min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)))
+    );
+}
+
+kernel void kernel_rope_norm_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src       = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_norm_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            float x0 = src[0];
+            float x1 = src[1];
+
+            dst_data[0] = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[1] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_neox_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    float theta_base = (float) pos[i2];
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src       = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections,
+        int  is_imrope
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global float * const src = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global float * dst_data  = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_multi_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections,
+        int  is_imrope
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1 + sections.s2 + sections.s3;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        if (i0 < n_dims) {
+            int ic = i0/2;
+
+            const int sector = (i0 / 2) % sect_dims;
+            float theta_base = 0.0f;
+
+            if (is_imrope) {
+                if (sector % 3 == 1 && sector < 3 * sections.s1) { // h
+                    theta_base = (float) pos[i2 + ne02 * 1];
+                } else if (sector % 3 == 2 && sector < 3 * sections.s2) { // w
+                    theta_base = (float) pos[i2 + ne02 * 2];
+                } else if (sector % 3 == 0 && sector < 3 * sections.s0) { // t
+                    theta_base = (float) pos[i2 + ne02 * 0];
+                } else { // e
+                    theta_base = (float) pos[i2 + ne02 * 3];
+                }
+            } else {
+                if (sector < sections.s0) {
+                    theta_base = pos[i2];
+                }
+                else if (sector >= sections.s0 && sector < sec_w) {
+                    theta_base = pos[i2 + ne2 * 1];
+                }
+                else if (sector >= sec_w && sector < sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 2];
+                }
+                else if (sector >= sec_w + sections.s2) {
+                    theta_base = pos[i2 + ne2 * 3];
+                }
+            }
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            float2 cos_sin_theta = rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+            global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+            dst_data[n_dims/2] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+        } else {
+            global half * const src = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            global half * dst_data  = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+kernel void kernel_rope_vision_f32(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global float * src      = (global float *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global float * dst_data = (global float *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
+
+kernel void kernel_rope_vision_f16(
+        global void * src0,
+        ulong offset0,
+        global int * src1,
+        ulong offset1,
+        global float * src2,
+        ulong offset2,
+        global half * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne0,
+        int ne1,
+        int ne2,
+        int ne3,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        int n_past,
+        int n_dims,
+        int n_ctx_orig,
+        float freq_base,
+        float freq_scale,
+        float ext_factor,
+        float attn_factor,
+        float beta_fast,
+        float beta_slow,
+        int4 sections
+) {
+    src0 = (global void*)((global char*)src0 + offset0);
+    src1 = (global int*)((global char*)src1 + offset1);
+    src2 = (global float*)((global char*)src2 + offset2);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int i3 = get_group_id(2);
+    int i2 = get_group_id(1);
+    int i1 = get_group_id(0);
+
+    float2 corr_dims = rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow);
+
+    global int * pos = src1;
+
+    const int sect_dims = sections.s0 + sections.s1;
+    const int sec_w = sections.s1 + sections.s0;
+
+    float inv_ndims = -1.f/n_dims;
+
+    for (int i0 = 2*get_local_id(0); i0 < ne0; i0 += 2*get_local_size(0)) {
+        int ic = i0/2;
+
+        const int sector = (i0/2) % sect_dims;
+        float theta_base = 0.0f;
+
+        if (sector < sections.s0) {
+            const int p = sector;
+            theta_base = pos[i2] * pow(freq_base, inv_ndims*2.0f*p);
+        } else if (sector >= sections.s0 && sector < sec_w) {
+            const int p = sector - sections.s0;
+            theta_base = pos[i2 + ne2] * pow(freq_base, inv_ndims*2.0f*p);
+        }
+
+        const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+        float2 cos_sin_theta = rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor);
+
+        global half * src      = (global half *)((global char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+        global half * dst_data = (global half *)((global char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+        const float x0 = src[0];
+        const float x1 = src[n_dims];
+
+        dst_data[0]      = x0*cos_sin_theta.s0 - x1*cos_sin_theta.s1;
+        dst_data[n_dims] = x0*cos_sin_theta.s1 + x1*cos_sin_theta.s0;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl
new file mode 100644
index 0000000..17ed97f
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/scale.cl
@@ -0,0 +1,27 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_scale_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        float scale,
+        float bias
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
+}
+
+kernel void kernel_scale_f32_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd,
+        float scale,
+        float bias
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+    dst[get_global_id(0)] = src0[get_global_id(0)] * scale + bias;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl
new file mode 100644
index 0000000..fc3ff7a
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/set_rows.cl
@@ -0,0 +1,208 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// v = { mp, L, d }
+inline uint fastdiv(uint n, uint4 v) {
+    uint msbs;
+    msbs = mul_hi(n, v.s0);
+    return (msbs + n) >> v.s1;
+}
+inline uint fastmod(uint n, uint4 v) {
+    uint q = fastdiv(n, v);
+    return n - q * v.s2;
+}
+
+kernel void kernel_set_rows_f32_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16_i64(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    long i1 = ((global long *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f32_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global float * dst_row = (global float *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = (float)src_row[ind];
+    }
+}
+
+kernel void kernel_set_rows_f16_i32(
+        global char * src0,
+        ulong         offset0,
+        global char * src1,
+        ulong         offset1,
+        global char * dst,
+        ulong         offsetd,
+        int           ne01,
+        ulong         nb01,
+        ulong         nb02,
+        ulong         nb03,
+        uint4         ne11,
+        uint4         ne12,
+        ulong         nb10,
+        ulong         nb11,
+        ulong         nb12,
+        int           nblk0,
+        ulong         nb1,
+        ulong         nb2,
+        ulong         nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0)*get_local_size(1) + get_local_id(1);
+
+    if (i01 >= ne01) {
+        return;
+    }
+
+    //int i12 = i03%ne12;
+    //int i11 = i02%ne11;
+    int i12 = fastmod(i03, ne12);
+    int i11 = fastmod(i02, ne11);
+
+    int i10 = i01;
+    int i1  = ((global int *)(src1 + i10*nb10 + i11*nb11 + i12*nb12))[0];
+
+    global half  * dst_row = (global half  *) (dst  +  i1*nb1  + i02*nb2  + i03*nb3);
+    global float * src_row = (global float *) (src0 + i01*nb01 + i02*nb02 + i03*nb03);
+
+    for (int ind = get_local_id(0); ind < nblk0; ind += get_local_size(0)) {
+        dst_row[ind] = src_row[ind];
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl
new file mode 100644
index 0000000..e3f669d
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sigmoid.cl
@@ -0,0 +1,29 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// sigmoid
+//------------------------------------------------------------------------------
+
+kernel void kernel_sigmoid_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)]));
+}
+
+kernel void kernel_sigmoid_f16(
+        global half * src0,
+        ulong offset0,
+        global half * dst,
+        ulong offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = 1.0f / (1.0f + exp(-src0[get_global_id(0)]));
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl
new file mode 100644
index 0000000..1d95e1b
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/silu.cl
@@ -0,0 +1,30 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// silu
+//------------------------------------------------------------------------------
+kernel void kernel_silu(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    float x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_silu_4(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * dst,
+        ulong offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    float4 x = src0[get_global_id(0)];
+    dst[get_global_id(0)] = x / (1.0f + exp(-x));
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
new file mode 100644
index 0000000..571d165
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f16.cl
@@ -0,0 +1,108 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half4  * pmask = src1 != src0 ? (global half4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + slope*(pmask ? convert_float4(pmask[i00]) : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
new file mode 100644
index 0000000..1f944b2
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_4_f32.cl
@@ -0,0 +1,108 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_4(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float4 * psrc4 = (global float4 *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float4 * pmask = src1 != src0 ? (global float4 *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float  * psrc2 = src2 != src0 ? (global float  *)(src2) : 0;
+    global float4 * pdst4 = (global float4 *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float lmax = fmax(fmax(lmax4.s0, lmax4.s1), fmax(lmax4.s2, lmax4.s3));
+
+    const float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+    float lsum = lsum4.s0 + lsum4.s1 + lsum4.s2 + lsum4.s3;
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        pdst4[i00] /= sum;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl
new file mode 100644
index 0000000..4baa6c2
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f16.cl
@@ -0,0 +1,107 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global half  * pmask = src1 != src0 ? (global half *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl
new file mode 100644
index 0000000..d503190
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softmax_f32.cl
@@ -0,0 +1,107 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_soft_max(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float scale,
+        float max_bias,
+        float m0,
+        float m1,
+        int n_head_log2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03%ne13;
+    int i12 = i02%ne12;
+    int i11 = i01;
+
+    global float * psrc0 = (global float *)(src0 + i01*nb01 + i02*nb02 + i03*nb03);
+    global float * pmask = src1 != src0 ? (global float *)(src1 + i11*nb11 + i12*nb12 + i13*nb13) : 0;
+    global float * psrc2 = src2 != src0 ? (global float *)(src2) : 0;
+    global float * pdst  = (global float *)(dst  + i01*nb1 + i02*nb2 + i03*nb3);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        int h = i02;
+
+        float base = h < n_head_log2 ? m0 : m1;
+        int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = psrc2 ? psrc2[i02] : -INFINITY;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        lmax = fmax(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+    float max = sub_group_reduce_max(lmax);
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max);
+        lsum += exp_psrc0;
+        // Remember the result of exp here. exp is expensive, so we really do not
+        // wish to compute it twice.
+        pdst[i00] = exp_psrc0;
+    }
+
+    float sum = sub_group_reduce_add(lsum);
+
+    if (psrc2) {
+        sum += exp(psrc2[i02] - max);
+    }
+
+    for (int i00 = get_local_id(0); i00 < ne00; i00 += get_local_size(0)) {
+        pdst[i00] /= sum;
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl
new file mode 100644
index 0000000..033766e
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/softplus.cl
@@ -0,0 +1,88 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// softplus
+//------------------------------------------------------------------------------
+inline float softplus_f32(float x){
+    float ax = fabs(x);
+    float m = fmax(x, 0.0f);
+    return log1p(exp(-ax)) + m;
+}
+
+kernel void kernel_softplus_f32_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = softplus_f32(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_softplus_f16_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = (half)(softplus_f32((float)(*src_val_ptr)));
+        }
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl
new file mode 100644
index 0000000..80745fc
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/solve_tri.cl
@@ -0,0 +1,51 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// solve_tri
+//------------------------------------------------------------------------------
+kernel void kernel_solve_tri_f32(
+        global uchar * src0,
+        ulong offset0,
+        global uchar * src1,
+        ulong offset1,
+        global uchar * dst,
+        ulong offsetd,
+        int n,
+        int k,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    int col = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    global const uchar * Lb = src0 + offset0 + i2 * nb02 + i3 * nb03;
+    global const uchar * Bb = src1 + offset1 + i2 * nb12 + i3 * nb13;
+    global       uchar * Xb = dst + offsetd + i2 * nb2 + i3 * nb3;
+
+    for(int row = 0; row < n; ++row){
+        global const float *pB = (global const float *)(Bb + row * nb11 + col * nb10);
+
+        float sum = 0.0f;
+        for(int j = 0; j < row; ++j){
+            global const float *pL = (global const float *)(Lb + row * nb01 + j * nb00);
+            global const float *pX = (global const float *)(Xb + j * nb1 + col * nb0);
+            sum += (*pL) * (*pX);
+        }
+
+        global const float * pDiag = (global const float *)(Lb + row * nb01 + row *nb00);
+        global float * pOut = (global float *)(Xb + row * nb1 + col *nb0);
+
+        *pOut = ((* pB) - sum) / (*pDiag);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl
new file mode 100644
index 0000000..4310906
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sqr.cl
@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqr_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
+
+kernel void kernel_sqr_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = src0[gid] * src0[gid];
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl
new file mode 100644
index 0000000..c59fbe0
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sqrt.cl
@@ -0,0 +1,53 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_sqrt_cont_f32(
+    global float * src0,
+    ulong          offset0,
+    global float * dst,
+    ulong          offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f32_4(
+    global float4 * src0,
+    ulong           offset0,
+    global float4 * dst,
+    ulong           offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = sqrt(src0[gid]);
+}
+
+kernel void kernel_sqrt_cont_f16(
+    global half * src0,
+    ulong         offset0,
+    global half * dst,
+    ulong         offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half(sqrt(convert_float(src0[gid])));
+}
+
+kernel void kernel_sqrt_cont_f16_4(
+    global half4 * src0,
+    ulong          offset0,
+    global half4 * dst,
+    ulong          offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    uint gid = get_global_id(0);
+    dst[gid] = convert_half4(sqrt(convert_float4(src0[gid])));
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl
new file mode 100644
index 0000000..7ae21ac
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/ssm_conv.cl
@@ -0,0 +1,77 @@
+kernel void kernel_ssm_conv_f32_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+){
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc  = ne10;
+
+    global float * s = (global float *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float * c = (global float *) (src1 + ir*nb11);
+    global float * d = (global float *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    d[0] = sumf;
+}
+
+kernel void kernel_ssm_conv_f32_f32_4(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    ulong         nb00,
+    ulong         nb01,
+    ulong         nb02,
+    int           ne10,
+    ulong         nb11,
+    ulong         nb0,
+    ulong         nb1,
+    ulong         nb2
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ir = get_global_id(0);
+    int i2 = get_global_id(1);
+    int i3 = get_global_id(2);
+
+    int nc = ne10;
+
+    global float4 * s = (global float4 *) (src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    global float4 * c = (global float4 *) (src1 + ir*nb11);
+    global float  * d = (global float  *) (dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int i0 = 0; i0 < nc/4; ++i0) {
+        sumf += dot(s[i0], c[i0]);
+    }
+
+    d[0] = sumf;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl
new file mode 100644
index 0000000..423ed59
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sub.cl
@@ -0,0 +1,138 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// div
+//------------------------------------------------------------------------------
+kernel void kernel_sub(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) - *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_sub_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] - src1[idx1];
+}
+
+kernel void kernel_sub_f16(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global half *)(dst_ptr + i0*nb0)) = *((global half *)(src0_ptr + i0*nb00)) - *((global half *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_sub_row_f16(
+        global half4 * src0,
+        ulong offset0,
+        global half4 * src1,
+        ulong offset1,
+        global half4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    src1 = (global half4*)((global char*)src1 + offset1);
+    dst = (global half4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] - src1[idx1];
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl
new file mode 100644
index 0000000..c5f7c57
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/sum_rows.cl
@@ -0,0 +1,39 @@
+
+kernel void kernel_sum_rows_f32(
+    global float *  src0,
+    ulong           offset0,
+    global float *  dst,
+    ulong           offsetd,
+    int             ne00,
+    int             ne01,
+    int             ne02,
+    int             ne03,
+    ulong           nb01,
+    ulong           nb02,
+    ulong           nb03,
+    ulong           nb1,
+    ulong           nb2,
+    ulong           nb3
+) {
+    src0 = (global float *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int i3 = get_global_id(2);
+    int i2 = get_global_id(1);
+    int i1 = get_global_id(0);
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    global float * src_row = (global float *) ((global char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    global float * dst_row = (global float *) ((global char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl
new file mode 100644
index 0000000..2c4887a
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tanh.cl
@@ -0,0 +1,109 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+kernel void kernel_tanh_f32(
+        global const float * src0,
+        ulong                offset0,
+        global       float * dst,
+        ulong                offsetd
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f32_4(
+        global const float4 * src0,
+        ulong                 offset0,
+        global       float4 * dst,
+        ulong                 offsetd
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    dst  = (global float4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f16(
+        global const half * src0,
+        ulong               offset0,
+        global       half * dst,
+        ulong               offsetd
+) {
+    src0 = (global half*)((global char*)src0 + offset0);
+    dst  = (global half*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f16_4(
+        global const half4 * src0,
+        ulong                offset0,
+        global       half4 * dst,
+        ulong                offsetd
+) {
+    src0 = (global half4*)((global char*)src0 + offset0);
+    dst  = (global half4*)((global char*)dst + offsetd);
+
+    dst[get_global_id(0)] = tanh(src0[get_global_id(0)]);
+}
+
+kernel void kernel_tanh_f32_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const float * x = (global const float *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       float * y = (global       float *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = tanh(*x);
+    }
+}
+
+kernel void kernel_tanh_f16_nc(
+        global const char * src0,
+        ulong               offset0,
+        global       char * dst,
+        ulong               offsetd,
+        int   ne00,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    dst  = dst + offsetd;
+
+    const int i3 = get_group_id(2);
+    const int i2 = get_group_id(1);
+    const int i1 = get_group_id(0);
+
+    for (int i0 = get_local_id(0); i0 < ne00; i0 += get_local_size(0)) {
+        global const half * x = (global const half *)(src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+        global       half * y = (global       half *)(dst  + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+        *y = tanh(*x);
+    }
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl
new file mode 100644
index 0000000..1279b65
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/transpose.cl
@@ -0,0 +1,117 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+// 16-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_16(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    half4 temp0 = read_imageh(input, (j_2+0)*cols+i);
+    half4 temp1 = read_imageh(input, (j_2+1)*cols+i);
+    half4 temp2 = read_imageh(input, (j_2+2)*cols+i);
+    half4 temp3 = read_imageh(input, (j_2+3)*cols+i);
+
+    write_imageh(output, (i_2+0)*rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imageh(output, (i_2+1)*rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}
+
+// Padded kernel for irregular shape
+kernel void kernel_transpose_16_4x1(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int j_2 = j << 2;
+
+    half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x;
+    half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x;
+    half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x;
+    half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x;
+
+    write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
+}
+
+// Transpose treating each element as 16-bit using buffer
+kernel void kernel_transpose_16_buf(
+    global const ushort * input,
+    global ushort * output,
+    const int ldi,
+    const int ldo
+) {
+    const int x = get_global_id(0);
+    const int y = get_global_id(1);
+
+    output[x*ldo + y] = input[y*ldi + x];
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+kernel void kernel_transpose_32(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+
+    float4 temp0 = read_imagef(input, (j_2+0)*cols+i);
+    float4 temp1 = read_imagef(input, (j_2+1)*cols+i);
+    float4 temp2 = read_imagef(input, (j_2+2)*cols+i);
+    float4 temp3 = read_imagef(input, (j_2+3)*cols+i);
+
+    write_imagef(output, (i_2+0)*rows+j, (float4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0));
+    write_imagef(output, (i_2+1)*rows+j, (float4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imagef(output, (i_2+2)*rows+j, (float4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imagef(output, (i_2+3)*rows+j, (float4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+
+}
+
+// 32-bit transpose, loading/storing a 4x4 tile of elements
+// Only used for activations
+// converts to FP16
+// also adds zero padding for non multiple of 8 prompt lengths
+kernel void kernel_transpose_32_16(__read_only image1d_buffer_t input, __write_only image1d_buffer_t output, const uint rows, const uint cols, const uint padded_rows) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int i_2 = i<<2;
+    const int j_2 = j<<2;
+    half4 temp0 = {0,0,0,0}; // initialize outputs to 0
+    half4 temp1 = {0,0,0,0};
+    half4 temp2 = {0,0,0,0};
+    half4 temp3 = {0,0,0,0};
+
+    if((j_2+0)*cols+i*4+3 < rows*cols*16){ // only load from a valid location. Otherwise keep register data as 0
+        temp0 = read_imageh(input, (j_2+0)*cols+i);
+    }
+    if((j_2+1)*cols+i*4+3 < rows*cols*16){
+        temp1 = read_imageh(input, (j_2+1)*cols+i);
+    }
+    if((j_2+2)*cols+i*4+3 < rows*cols*16){
+        temp2 = read_imageh(input, (j_2+2)*cols+i);
+    }
+    if((j_2+3)*cols+i*4+3 < rows*cols*16){
+        temp3 = read_imageh(input, (j_2+3)*cols+i);
+    }
+
+    write_imageh(output, (i_2+0)*padded_rows+j, (half4)(temp0.s0, temp1.s0, temp2.s0, temp3.s0)); // no conditionals for output, includes zero padding
+    write_imageh(output, (i_2+1)*padded_rows+j, (half4)(temp0.s1, temp1.s1, temp2.s1, temp3.s1));
+    write_imageh(output, (i_2+2)*padded_rows+j, (half4)(temp0.s2, temp1.s2, temp2.s2, temp3.s2));
+    write_imageh(output, (i_2+3)*padded_rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl
new file mode 100644
index 0000000..35cdd54
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tri.cl
@@ -0,0 +1,32 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// tri
+//------------------------------------------------------------------------------
+__kernel void kernel_tri_f32(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int n,
+        int ne0,
+        int ne1,
+        int tri_type
+) {
+    src0 = (global float*)((global char*)src0 + offset0);
+    dst = (global float*)((global char*)dst + offsetd);
+
+    int idx = get_global_id(0);
+    if (idx >= n) return;
+
+    int i0 = idx % ne0;
+    int i1 = (idx / ne0) % ne1;
+
+    int keep = 0;
+    if (tri_type == 0) keep = (i0 >= i1);
+    else if (tri_type == 1) keep = (i0 >  i1);
+    else if (tri_type == 2) keep = (i0 <= i1);
+    else                    keep = (i0 <  i1);
+
+    dst[idx] = keep ? src0[idx] : 0.0f;
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl
new file mode 100644
index 0000000..21444bd
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/tsembd.cl
@@ -0,0 +1,48 @@
+kernel void kernel_timestep_embedding(
+    global const void * p_timesteps,
+    ulong off_timesteps,
+    global void * p_dst,
+    ulong off_dst,
+    int dst_nb1_bytes,
+    int logical_dim,
+    int max_period
+) {
+    int local_i;
+    int local_j;
+    int local_half_dim;
+    float local_timestep_val;
+    float local_freq;
+    float local_arg;
+    global float * local_embed_data_ptr;
+    global const float * local_timesteps_input_ptr;
+    global float * local_dst_output_base_ptr;
+
+    local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps);
+    local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst);
+
+    local_i = get_global_id(1);
+    local_j = get_global_id(0);
+
+    local_half_dim = logical_dim / 2;
+    local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
+
+    if (logical_dim % 2 != 0 && local_j == local_half_dim) {
+        local_embed_data_ptr[2 * local_half_dim] = 0.0f;
+    }
+
+    if (local_j >= local_half_dim) {
+        return;
+    }
+
+    local_timestep_val = local_timesteps_input_ptr[local_i];
+
+    if (local_half_dim == 0) {
+        local_freq = 1.0f;
+    } else {
+        local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim);
+    }
+
+    local_arg = local_timestep_val * local_freq;
+    local_embed_data_ptr[local_j] = cos(local_arg);
+    local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg);
+}
diff --git a/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl b/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl
new file mode 100644
index 0000000..25c6835
--- /dev/null
+++ b/llama.cpp/ggml/src/ggml-opencl/kernels/upscale.cl
@@ -0,0 +1,120 @@
+kernel void kernel_upscale(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne10,
+    int ne11,
+    int ne12,
+    int ne13,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10 * ne11 * ne12 * ne13;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = index / (ne10 * ne11 * ne12);
+
+    int i00 = (int)(i10 / sf0);
+    int i01 = (int)(i11 / sf1);
+    int i02 = (int)(i12 / sf2);
+    int i03 = (int)(i13 / sf3);
+
+    ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00;
+    global const float * src_element_ptr = (global const float *)(src_base + offset_src_element);
+
+    dst_base[index] = *src_element_ptr;
+}
+
+kernel void kernel_upscale_bilinear(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne00_src,
+    int ne01_src,
+    int ne10_dst,
+    int ne11_dst,
+    int ne12_dst,
+    int ne13_dst,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3,
+    float pixel_offset
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10_dst = index % ne10_dst;
+    int i11_dst = (index / ne10_dst) % ne11_dst;
+    int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    int i02_src = (int)(i12_dst / sf2);
+    int i03_src = (int)(i13_dst / sf3);
+
+    float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    long y0_src = (long)floor(y_src_f);
+    long y1_src = y0_src + 1;
+
+    y0_src = max(0L, min(y0_src, (long)ne01_src - 1));
+    y1_src = max(0L, min(y1_src, (long)ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    long x0_src = (long)floor(x_src_f);
+    long x1_src = x0_src + 1;
+
+    x0_src = max(0L, min(x0_src, (long)ne00_src - 1));
+    x1_src = max(0L, min(x1_src, (long)ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst_base[index] = result;
+}