ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

Fix mla = 0 (#1130)

Browse Source 
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2026-01-10 10:34:30 +02:00
committed by GitHub
parent c7dba35702
commit c7348f6f55
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ggml/src/ggml-cuda/fattn.cu
View File

@@ -94,7 +94,7 @@ void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst
//const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16; //const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
//const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < CC_ADA_LOVELACE && !mma_needs_data_conversion; //const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < CC_ADA_LOVELACE && !mma_needs_data_conversion;
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !(Q->ne[1] == 1 && n_swa > 0 && K->ne[0] == V->ne[0]); const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !(Q->ne[1] == 1 && n_swa > 0 && K->ne[0] == V->ne[0]);
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*WARP_SIZE) == 0; const bool can_use_vector_kernel = Q->ne[0] <= 256 && K->ne[0] == V->ne[0] && Q->ne[0] % (2*WARP_SIZE) == 0;
if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1 && !ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) { if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1 && !ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
ggml_cuda_flash_attn_ext_vec_f32(ctx, dst); ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
return; return;
@@ -172,7 +172,7 @@ bool ggml_cuda_fattn_is_supported(ggml_backend_cuda_context & ctx, const ggml_te
//const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16; //const bool mma_needs_data_conversion = K->type != GGML_TYPE_F16 || V->type != GGML_TYPE_F16;
//const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < CC_ADA_LOVELACE && !mma_needs_data_conversion; //const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && cc < CC_ADA_LOVELACE && !mma_needs_data_conversion;
const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !(Q->ne[1] == 1 && n_swa > 0 && K->ne[0] == V->ne[0]); const bool mma_faster_for_bs1 = new_mma_available(cc) && gqa_opt_applies && !(Q->ne[1] == 1 && n_swa > 0 && K->ne[0] == V->ne[0]);
const bool can_use_vector_kernel = Q->ne[0] <= 256 && Q->ne[0] % (2*WARP_SIZE) == 0; const bool can_use_vector_kernel = Q->ne[0] <= 256 && K->ne[0] == V->ne[0] && Q->ne[0] % (2*WARP_SIZE) == 0;
if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1 && !ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) { if (Q->ne[1] == 1 && can_use_vector_kernel && !mma_faster_for_bs1 && !ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
return ggml_cuda_fattn_vec_f32_is_supported(ctx, dst); return ggml_cuda_fattn_vec_f32_is_supported(ctx, dst);
} }