From aff7aa0cf6e38d2d8ceca918c2c5442e2c0e225d Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Sun, 25 Jan 2026 06:52:04 +0000 Subject: [PATCH] Add condition --- ggml/src/ggml-cuda/fattn-mma-f16.cu | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-cuda/fattn-mma-f16.cu b/ggml/src/ggml-cuda/fattn-mma-f16.cu index 2c36c1c0..07bdabd3 100644 --- a/ggml/src/ggml-cuda/fattn-mma-f16.cu +++ b/ggml/src/ggml-cuda/fattn-mma-f16.cu @@ -217,7 +217,8 @@ void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, ggml_tens GGML_ASSERT(Q->ne[2] % K->ne[2] == 0); const int gqa_ratio = Q->ne[2] / K->ne[2]; - if (gqa_ratio == 12) { + if (gqa_ratio == 12 && Q->ne[1] == 1 && K->ne[1]*K->ne[2] >= 65536) { + // This is a hack to improve GLM-4.5/4.6/4.7/AIR TG performance glm45_flash_attention(ctx, dst); return; }