Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml-cuda: native bf16 flash attention for vec kernel (#20525)

Browse Source 
* ggml-cuda: native bf16 flash attention for vec and tile kernels

mma kernel still converts bf16 to fp16 before launch, native mma bf16 todo

* ggml-cuda: address code owner review feedback

reverted tile kernel changes to avoid larger refactor

* fix ci failures on turing and hip

* fix bf16 vec kernel compile on hip v_dot2 platforms

* add comments

---------

Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
This commit is contained in:
Patrick Buckley
2026-03-22 11:05:51 +01:00
committed by GitHub
co-authored by Johannes Gäßler
parent ccb87fa3ee
commit db9d8aa428
21 changed files with 197 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn.cu
+16
View File
@@ -224,6 +224,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_F16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_F16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_F16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_F16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
@@ -231,6 +232,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_Q4_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
@@ -238,6 +240,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q4_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_Q5_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
@@ -245,6 +248,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_Q5_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
@@ -252,6 +256,7 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q5_1)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
@@ -259,10 +264,20 @@ static void ggml_cuda_flash_attn_ext_vec(ggml_backend_cuda_context & ctx, ggml_t
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_1, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_0, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q5_1, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_BF16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
#else
FATTN_VEC_CASES_ALL_D(GGML_TYPE_F16, GGML_TYPE_F16)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
FATTN_VEC_CASES_ALL_D(GGML_TYPE_BF16, GGML_TYPE_BF16)
#endif // GGML_CUDA_FA_ALL_QUANTS
GGML_ABORT("fatal error");
@@ -355,6 +370,7 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
#endif // GGML_CUDA_FA_ALL_QUANTS
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q8_0:
case GGML_TYPE_BF16:
break;
default:
return BEST_FATTN_KERNEL_NONE;