Commit 77e2db8ca for llama.cpp
commit 77e2db8cabdb83f8da59d8584a381b2299d8a171
Author: Jiří Svítil <32205472+mistrjirka@users.noreply.github.com>
Date: Wed Nov 11 09:45:34 2026 +0100
cuda : add sm70 FlashAttention config case for DKQ=256, DV=256, ncols=64 (#27997)
* CUDA : add sm70 FlashAttention config case for the 256x256 head
* CUDA : tune sm70 FlashAttention for 256x256 ncols=32
Assisted-by: Codex
---------
Co-authored-by: Jirka Svítil <mistrjirka@users.noreply.github.com>
diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
index 2ce2fb261..07a6db842 100644
--- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh
+++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh
@@ -133,6 +133,9 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2, 32, 160, 128, 64, 1, false);
GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1, 32, 160, 128, 64, 1, false);
+ GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2, 32, 128, 128, 64, 1, false);
+ GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2, 32, 128, 64, 128, 2, false);
+
// TODO tune specifically for Volta
return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
}