From 7aaed3f48815df1af57ce0d1d0a0516c29e32c16 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Wed, 28 Jan 2026 12:37:51 +0000 Subject: [PATCH] Apply offfset to KQ_max in CUDA flash attention --- ggml/src/ggml-cuda/fattn-mma-f16.cuh | 4 ++-- ggml/src/ggml-cuda/fattn-new-mma.cu | 4 ++-- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cuh b/ggml/src/ggml-cuda/fattn-mma-f16.cuh index 27328319..050186ce 100644 --- a/ggml/src/ggml-cuda/fattn-mma-f16.cuh +++ b/ggml/src/ggml-cuda/fattn-mma-f16.cuh @@ -264,7 +264,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( for (int k = 0; k < KQ_per_iter/(np*tile_C_KQ::I); ++k) { #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { - KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l]); + KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l] + FATTN_KQ_MAX_OFFSET); } } @@ -319,7 +319,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( #pragma unroll for (int l = 0; l < tile_C_KQ_16::ne; ++l) { const int KQ_index = 2*t + (l/2) % 2; - KQ_max_new[KQ_index] = fmaxf(KQ_max_new[KQ_index], KQ_C_16[k*ntiles/2 + t].x[l]); + KQ_max_new[KQ_index] = fmaxf(KQ_max_new[KQ_index], KQ_C_16[k*ntiles/2 + t].x[l] + FATTN_KQ_MAX_OFFSET); } } } diff --git a/ggml/src/ggml-cuda/fattn-new-mma.cu b/ggml/src/ggml-cuda/fattn-new-mma.cu index 52a14639..89ee392e 100644 --- a/ggml/src/ggml-cuda/fattn-new-mma.cu +++ b/ggml/src/ggml-cuda/fattn-new-mma.cu @@ -702,7 +702,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( for (int k = 0; k < c::nbatch_fa/(np*tile_C_KQ::I); ++k) { #pragma unroll for (int l = 0; l < tile_C_KQ::ne; ++l) { - KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l]); + KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k].x[l] + FATTN_KQ_MAX_OFFSET); } } @@ -756,7 +756,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter( #pragma unroll for (int l = 0; l < tile_C_KQ_16::ne; ++l) { const int KQ_index = 2*t + (l/2) % 2; - KQ_max_new[KQ_index] = fmaxf(KQ_max_new[KQ_index], KQ_C_16[k*ntiles/2 + t].x[l]); + KQ_max_new[KQ_index] = fmaxf(KQ_max_new[KQ_index], KQ_C_16[k*ntiles/2 + t].x[l] + FATTN_KQ_MAX_OFFSET); } } }