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: FA support for Deepseek (Ampere or newer) (#13306)

Browse Source 
* CUDA: FA support for Deepseek (Ampere or newer)

* do loop unrolling via C++ template
This commit is contained in:
Johannes Gäßler
2025-05-09 13:34:58 +02:00
committed by GitHub
parent 27ebfcacba
commit 0cf6725e9f
33 changed files with 852 additions and 547 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-cuda/ggml-cuda.cu
View File

@@ -3215,16 +3215,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
return false;
#endif // FLASH_ATTN_AVAILABLE
if (op->src[1]->ne[0] != op->src[2]->ne[0]) {
// different head sizes of K and V are not supported yet
return false;
const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
if (!new_mma_available(cc) || cc < GGML_CUDA_CC_AMPERE) {
return false;
}
const int gqa_ratio = op->src[0]->ne[2] / op->src[1]->ne[2];
return op->src[1]->ne[0] == 576 && op->src[2]->ne[0] == 512 && op->src[3] && gqa_ratio % 16 == 0;
}
if (op->src[0]->ne[0] == 192) {
return false;
}
if (op->src[0]->ne[0] == 576) {
// DeepSeek MLA
return false;
}
if (op->src[0]->ne[3] != 1) {
return false;
}