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

CUDA: fix Gemma E4B MTP FlashAttention (#25148)

Browse Source 
* CUDA: fix Gemma E4B MTP FlashAttention

* remove unused template declaration
This commit is contained in:
Johannes Gäßler
2026-06-30 14:06:54 +02:00
committed by GitHub
parent f708a5b2ca
commit e495d1e748
8 changed files with 23 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/fattn-tile.cuh
+9 -5
View File
@@ -76,6 +76,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
@@ -144,6 +145,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_nv
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 16, 256, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 32, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 32, 64)
@@ -219,6 +221,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 512, 1, 128, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 2, 64, 64)
@@ -296,6 +299,7 @@ static constexpr __host__ __device__ uint32_t ggml_cuda_fattn_tile_get_config_am
GGML_CUDA_FATTN_TILE_CONFIG_CASE(320, 256, 32, 256, 2, 128, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 2, 64, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 4, 128, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 8, 256, 2, 64, 64)
GGML_CUDA_FATTN_TILE_CONFIG_CASE(512, 512, 16, 256, 4, 64, 64)
@@ -1308,12 +1312,12 @@ static void launch_fattn_tile_switch_ncols2(ggml_backend_cuda_context & ctx, ggm
return;
}
if constexpr (DV <= 256) {
if (use_gqa_opt && gqa_ratio % 2 == 0) {
launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
return;
}
if (use_gqa_opt && gqa_ratio % 2 == 0) {
launch_fattn_tile_switch_ncols1<DKQ, DV, 2, use_logit_softcap>(ctx, dst);
return;
}
if constexpr (DV <= 256) {
launch_fattn_tile_switch_ncols1<DKQ, DV, 1, use_logit_softcap>(ctx, dst);
return;
}