1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
|
#include "common.cuh"
#include "wkv.cuh"
template <int block_size>
static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
const int tid = threadIdx.x;
const int bid = blockIdx.x;
const int head_size = block_size;
const int batch_i = bid / H;
const int head_i = bid % H;
const int state_size = C * head_size;
const int n_seq_tokens = T / B;
float state[head_size];
__shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
#pragma unroll
for (int i = 0; i < head_size; i++) {
state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
}
__syncthreads();
_tf[tid] = tf[head_i * head_size + tid];
__syncthreads();
for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
__syncthreads();
_k[tid] = k[t];
_r[tid] = r[t];
_td[tid] = td[t];
__syncthreads();
const float _v = v[t];
float y = 0;
for (int j = 0; j < head_size; j += 4) {
const float4& k = (float4&)(_k[j]);
const float4& r = (float4&)(_r[j]);
const float4& tf = (float4&)(_tf[j]);
const float4& td = (float4&)(_td[j]);
float4& s = (float4&)(state[j]);
float4 kv;
kv.x = k.x * _v;
kv.y = k.y * _v;
kv.z = k.z * _v;
kv.w = k.w * _v;
y += r.x * (tf.x * kv.x + s.x);
y += r.y * (tf.y * kv.y + s.y);
y += r.z * (tf.z * kv.z + s.z);
y += r.w * (tf.w * kv.w + s.w);
s.x = s.x * td.x + kv.x;
s.y = s.y * td.y + kv.y;
s.z = s.z * td.z + kv.z;
s.w = s.w * td.w + kv.w;
}
dst[t] = y;
}
#pragma unroll
for (int i = 0; i < head_size; i++) {
dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
}
}
template <int block_size>
static __global__ void rwkv_wkv7_f32(const int B, const int T, const int C, const int H, const float * r, const float * w, const float * k, const float * v, const float * a, const float * b, const float * s, float * dst) {
const int tid = threadIdx.x;
const int bid = blockIdx.x;
const int head_size = block_size;
const int batch_i = bid / H;
const int head_i = bid % H;
const int state_size = C * head_size;
const int n_seq_tokens = T / B;
float state[head_size];
__shared__ float _r[head_size], _w[head_size], _k[head_size], _a[head_size], _b[head_size];
#ifndef GGML_USE_MUSA
#pragma unroll
#endif
for (int i = 0; i < head_size; i++) {
state[i] = s[batch_i * state_size + head_i * head_size * head_size + tid * head_size + i];
}
for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
__syncthreads();
_r[tid] = r[t];
_w[tid] = w[t];
_k[tid] = k[t];
_a[tid] = a[t];
_b[tid] = b[t];
__syncthreads();
float sa = 0;
#pragma unroll
for (int j = 0; j < head_size; j += 4)
{
const float4& a = (float4&)(_a[j]);
const float4& s = (float4&)(state[j]);
sa += a.x * s.x;
sa += a.y * s.y;
sa += a.z * s.z;
sa += a.w * s.w;
}
const float _v = v[t];
float y = 0;
for (int j = 0; j < head_size; j += 4) {
const float4& r = (float4&)(_r[j]);
const float4& w = (float4&)(_w[j]);
const float4& k = (float4&)(_k[j]);
const float4& b = (float4&)(_b[j]);
float4& s = (float4&)(state[j]);
float4 kv;
kv.x = k.x * _v;
kv.y = k.y * _v;
kv.z = k.z * _v;
kv.w = k.w * _v;
s.x = s.x * w.x + kv.x + sa * b.x;
s.y = s.y * w.y + kv.y + sa * b.y;
s.z = s.z * w.z + kv.z + sa * b.z;
s.w = s.w * w.w + kv.w + sa * b.w;
y += s.x * r.x;
y += s.y * r.y;
y += s.z * r.z;
y += s.w * r.w;
}
dst[t] = y;
}
#pragma unroll
for (int i = 0; i < head_size; i++) {
dst[T * C + batch_i * state_size + head_i * head_size * head_size + tid * head_size + i] = state[i];
}
}
void ggml_cuda_op_rwkv_wkv6(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const float * k_d = (const float *)dst->src[0]->data;
const float * v_d = (const float *)dst->src[1]->data;
const float * r_d = (const float *)dst->src[2]->data;
const float * tf_d = (const float *)dst->src[3]->data;
const float * td_d = (const float *)dst->src[4]->data;
const float * s_d = (const float *)dst->src[5]->data;
const int64_t B = dst->src[5]->ne[1];
const int64_t T = dst->src[0]->ne[2];
const int64_t C = dst->ne[0];
const int64_t H = dst->src[0]->ne[1];
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
GGML_ASSERT(C % H == 0);
GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
if (C / H == CUDA_WKV_BLOCK_SIZE) {
rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
} else {
rwkv_wkv_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
}
}
void ggml_cuda_op_rwkv_wkv7(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
const float * r_d = (const float *)dst->src[0]->data;
const float * w_d = (const float *)dst->src[1]->data;
const float * k_d = (const float *)dst->src[2]->data;
const float * v_d = (const float *)dst->src[3]->data;
const float * a_d = (const float *)dst->src[4]->data;
const float * b_d = (const float *)dst->src[5]->data;
const float * s_d = (const float *)dst->src[6]->data;
const int64_t B = dst->src[6]->ne[1];
const int64_t T = dst->src[0]->ne[2];
const int64_t C = dst->ne[0];
const int64_t H = dst->src[0]->ne[1];
float * dst_d = (float *)dst->data;
cudaStream_t stream = ctx.stream();
GGML_ASSERT(dst->src[6]->type == GGML_TYPE_F32);
GGML_ASSERT(C % H == 0);
GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE || C / H == CUDA_WKV_BLOCK_SIZE * 2);
if (C / H == CUDA_WKV_BLOCK_SIZE) {
rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
} else {
rwkv_wkv7_f32<CUDA_WKV_BLOCK_SIZE * 2><<<B * H, C / H, 0, stream>>>(B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d);
}
}
|
