llama.cpp-sm70-cuda12.9 / sm70-flash-attention.patch
orwelian84's picture
llama.cpp f74e2fa for NVIDIA Volta (sm_70), CUDA 12.9 — released 2026-10-20 with the Q5_1 caveat stated
0b864c5 verified
Raw History Blame Contribute Delete
417 Bytes
--- a/ggml/src/ggml-cuda/fattn.cu
+++ b/ggml/src/ggml-cuda/fattn.cu
@@ -175,7 +175,8 @@
}
if constexpr (ncols2 <= 16) {
- if (Q->ne[1] <= 16/ncols2) {
+ // Volta MMA kernels require at least 32 columns.
+ if (!volta_mma_available(cc) && Q->ne[1] <= 16/ncols2) {
ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
return;
}