ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-13 07:20:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

Make biased gemv fusion optional (#931)

Browse Source 
* Make biased gemv fusion optional

* Fix one path through gemv fusion

* Remove forgotten printf

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-11-09 19:09:47 +02:00
committed by GitHub
parent 7df9947923
commit 56ee303254
1 changed files with 12 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/ggml-cuda.cu
View File

@@ -2067,6 +2067,8 @@ static int ggml_cuda_mul_mat_q(ggml_backend_cuda_context & ctx, const ggml_tenso
auto stream = ctx.stream();
auto fusion = ctx.fusion;
auto ne10_padded = GGML_PAD(src1->ne[0], MATRIX_ROW_PADDING);
auto nb10_padded = ne10_padded*sizeof(block_q8_1)/QK8_1;
auto quantized_size = nb10_padded*ggml_nrows(src1);
@@ -2081,7 +2083,7 @@ static int ggml_cuda_mul_mat_q(ggml_backend_cuda_context & ctx, const ggml_tenso
// The code below handles the case when Q, K, V have a bias applied after the resepctive matrix multiplication.
// In that case the graph contains mul_mat(Q) -> mul_mat(K) -> mul_mat(V) -> add(Q) -> add(K) -> add(V)
if (cgraph && node_n + 5 < cgraph->n_nodes &&
if (fusion && cgraph && node_n + 5 < cgraph->n_nodes &&
cgraph->nodes[node_n+1]->op == GGML_OP_MUL_MAT &&
cgraph->nodes[node_n+2]->op == GGML_OP_MUL_MAT &&
ggml_is_quantized(cgraph->nodes[node_n+1]->src[0]->type) &&
@@ -2100,7 +2102,7 @@ static int ggml_cuda_mul_mat_q(ggml_backend_cuda_context & ctx, const ggml_tenso
CUDA_CHECK(cudaGetLastError());
}
node_n += 5;
} else if (cgraph && node_n + 1 < cgraph->n_nodes &&
} else if (fusion && cgraph && node_n + 1 < cgraph->n_nodes &&
cgraph->nodes[node_n+1]->op == GGML_OP_ADD &&
dst == cgraph->nodes[node_n+1]->src[0] &&
dst->ne[0] == cgraph->nodes[node_n+1]->src[1]->ne[0] &&
@@ -2136,15 +2138,15 @@ static int ggml_cuda_mul_mat_q(ggml_backend_cuda_context & ctx, const ggml_tenso
if (dst->op != GGML_OP_MUL_MAT || dst->src[1] != src1 || !ggml_is_quantized(dst->src[0]->type)) break;
if (!is_gemv && mmq_get_q8_1_ds_layout(src0->type) != mmq_get_q8_1_ds_layout(dst->src[0]->type)) break;
if (is_gemv) {
if (node_n + 1 < cgraph->n_nodes &&
cgraph->nodes[node_n+1]->op == GGML_OP_ADD &&
dst == cgraph->nodes[node_n+1]->src[0] &&
dst->ne[0] == cgraph->nodes[node_n+1]->src[1]->ne[0] &&
cgraph->nodes[node_n+1]->src[1]->type == GGML_TYPE_F32 &&
ggml_nrows(cgraph->nodes[node_n+1]->src[1]) == 1) {
if (fusion && node_n + 2 < cgraph->n_nodes &&
cgraph->nodes[node_n+2]->op == GGML_OP_ADD &&
dst == cgraph->nodes[node_n+2]->src[0] &&
dst->ne[0] == cgraph->nodes[node_n+2]->src[1]->ne[0] &&
cgraph->nodes[node_n+2]->src[1]->type == GGML_TYPE_F32 &&
ggml_nrows(cgraph->nodes[node_n+2]->src[1]) == 1) {
// We have a bias applied after the matrix multiplication and we can fuse it
ggml_cuda_op_mul_mat_vec_q_biased(ctx, dst->src[0], src1, cgraph->nodes[node_n+1], cgraph->nodes[node_n+1]->src[1],
(const char *)dst->src[0]->data, nullptr, src1_quantized.get(), (float *)cgraph->nodes[node_n+1]->data,
ggml_cuda_op_mul_mat_vec_q_biased(ctx, dst->src[0], src1, cgraph->nodes[node_n+2], cgraph->nodes[node_n+2]->src[1],
(const char *)dst->src[0]->data, nullptr, src1_quantized.get(), (float *)cgraph->nodes[node_n+2]->data,
0, dst->src[0]->ne[1], src1->ne[1], ne10_padded, stream);
++node_n;
} else {