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/template-instances/fattn-mma-f16-instance-ncols1_16-ncols2_2.cu
+1
View File
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 16, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 16, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 16, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 16, 2);
DECL_FATTN_MMA_F16_CASE(512, 512, 16, 2);
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_32-ncols2_2.cu
+1
View File
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 32, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 32, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 32, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 32, 2);
DECL_FATTN_MMA_F16_CASE(512, 512, 32, 2);
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_4-ncols2_2.cu
+1
View File
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 4, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 4, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 4, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 4, 2);
DECL_FATTN_MMA_F16_CASE(512, 512, 4, 2);
ggml/src/ggml-cuda/template-instances/fattn-mma-f16-instance-ncols1_8-ncols2_2.cu
+1
View File
@@ -8,3 +8,4 @@ DECL_FATTN_MMA_F16_CASE(96, 96, 8, 2);
DECL_FATTN_MMA_F16_CASE(112, 112, 8, 2);
DECL_FATTN_MMA_F16_CASE(128, 128, 8, 2);
DECL_FATTN_MMA_F16_CASE(256, 256, 8, 2);
DECL_FATTN_MMA_F16_CASE(512, 512, 8, 2);
ggml/src/ggml-cuda/template-instances/generate_cu_files.py
+1 -1
View File
@@ -92,7 +92,7 @@ for ncols in [8, 16, 32, 64]:
continue
if head_size_kq == 320 and ncols2 != 32: # Mistral Small 4
continue
if head_size_kq == 512 and ncols2 not in (4, 8): # Gemma 4
if head_size_kq == 512 and ncols2 not in (2, 4, 8): # Gemma 4 (+ MTP)
continue
if head_size_kq == 576 and ncols2 not in (4, 16, 32): # Deepseek, GLM 4.7 Flash
continue