From b333b06772c89d96aacb5490d6a219fba7c09cc6 Mon Sep 17 00:00:00 2001
From: Mitja Felicijan <mitja.felicijan@gmail.com>
Date: Thu, 12 Feb 2026 20:57:17 +0100
Subject: Engage!

---
 llama.cpp/src/llama-adapter.h | 86 +++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 86 insertions(+)
 create mode 100644 llama.cpp/src/llama-adapter.h

(limited to 'llama.cpp/src/llama-adapter.h')

diff --git a/llama.cpp/src/llama-adapter.h b/llama.cpp/src/llama-adapter.h
new file mode 100644
index 0000000..d275d25
--- /dev/null
+++ b/llama.cpp/src/llama-adapter.h
@@ -0,0 +1,86 @@
+#pragma once
+
+#include "llama.h"
+
+#include "ggml-cpp.h"
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+// TODO: pimpl
+
+//
+// llama_adapter_cvec
+//
+
+struct llama_adapter_cvec {
+    ggml_tensor * tensor_for(int il) const;
+
+    ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int  il) const;
+
+    bool apply(
+            const llama_model & model,
+            const float * data,
+            size_t len,
+            int32_t n_embd,
+            int32_t il_start,
+            int32_t il_end);
+
+private:
+    bool init(const llama_model & model);
+
+    int32_t layer_start = -1;
+    int32_t layer_end   = -1;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    std::vector<ggml_tensor *> tensors; // per layer
+};
+
+//
+// llama_adapter_lora
+//
+
+struct llama_adapter_lora_weight {
+    ggml_tensor * a = nullptr;
+    ggml_tensor * b = nullptr;
+
+    // get actual scale based on rank and alpha
+    float get_scale(float alpha, float adapter_scale) const {
+        const float rank  = (float) b->ne[0];
+        const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
+        return scale;
+    }
+
+    llama_adapter_lora_weight() = default;
+    llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
+};
+
+struct llama_adapter_lora {
+    // map tensor name to lora_a_b
+    std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
+
+    std::vector<ggml_context_ptr> ctxs;
+    std::vector<ggml_backend_buffer_ptr> bufs;
+
+    float alpha;
+
+    // gguf metadata
+    std::unordered_map<std::string, std::string> gguf_kv;
+
+    // activated lora (aLoRA)
+    std::vector<llama_token> alora_invocation_tokens;
+
+    llama_adapter_lora() = default;
+    ~llama_adapter_lora() = default;
+
+    llama_adapter_lora_weight * get_weight(ggml_tensor * w);
+
+    uint32_t get_n_nodes() const {
+        return ab_map.size() * 6u; // a, b, scale, add, 2 x mul_mat
+    }
+};
+
+using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;
-- 
cgit v1.2.3