happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fixes

This commit is contained in:
Ruben Ortlam 2026-02-12 13:58:44 +01:00
parent 3946eb657f
commit cd54ba2b86
1 changed files with 6 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -3256,18 +3256,19 @@ static void ggml_vk_load_shaders(vk_device& device) {
bool aligned = fa.first.aligned; \
bool f32acc = fa.first.f32acc; \
uint32_t flags = fa.first.flags; \
bool fa_ds = path == FA_SCALAR && disable_subgroups; \
if (path == FAPATH) { \
if (aligned) { \
if (f32acc) { \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,rows,small_cache), true, (!disable_subgroups && (FAPATH!=FA_COOPMAT2)), ((!disable_subgroups && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,rows,small_cache), true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
} else { \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,rows,small_cache), true, (!disable_subgroups && (FAPATH!=FA_COOPMAT2)), ((!disable_subgroups && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_aligned_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,0,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,0,TYPE,rows,small_cache,flags), fa_align(FAPATH,HSK,HSV,TYPE,rows,small_cache), true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
} \
} else { \
if (f32acc) { \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,rows,small_cache,flags), 1, true, (!disable_subgroups && (FAPATH!=FA_COOPMAT2)), ((!disable_subgroups && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f32acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,rows,small_cache,flags), 1, true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
} else { \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,rows,small_cache,flags), 1, true, (!disable_subgroups && (FAPATH!=FA_COOPMAT2)), ((!disable_subgroups && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
ggml_vk_create_pipeline(device, fa.second, "flash_attn_f32_f16_f16acc" #NAMELC, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _len, flash_attn_f32_f16_ ## NAMELC ## _f16acc ## SUFFIX ## _data, "main", 7, sizeof(vk_flash_attn_push_constants), fa_wg_denoms(FAPATH, HSK,HSV,1,TYPE,rows,small_cache), fa_spec_constants(FAPATH, HSK,HSV,1,TYPE,rows,small_cache,flags), 1, true, (!fa_ds && (FAPATH!=FA_COOPMAT2)), ((!fa_ds && (FAPATH!=FA_COOPMAT2)) ? device->subgroup_size : 0)); \
} \
} \
} \
@ -8596,7 +8597,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
FaRows rows;
if (N == 1) {
rows = FA_ROWS_1;
} else if (N <= 8) {
} else if (N <= (path == FA_COOPMAT2 ? flash_attention_num_small_rows : 8)) {
rows = FA_ROWS_SMALL;
} else {
rows = FA_ROWS_LARGE;