CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

CUDA: attention sinks for mma FlashAttention (#15157)

Browse Source
This commit is contained in:
Johannes Gäßler
2025-08-08 08:19:58 +02:00
committed by GitHub
parent aaa3d07ae7
commit 1425f587a8
3 changed files with 73 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -3532,7 +3532,8 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
}
// TODO: more general-purpose attention sink support [TAG_ATTN_SINKS]
if (op->src[4] && op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) { // currently only sinks for head_size 64 and 128 are supported
if (op->src[4] && !fp16_mma_available(ggml_cuda_info().devices[dev_ctx->device].cc)
&& op->src[0]->ne[0] != 64 && op->src[0]->ne[0] != 128) {
return false;
}
if (op->src[0]->ne[0] == 192) {