From ef64b1a1710a32f63e7673d34263c2e8f24eaed3 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Mon, 10 Nov 2025 07:43:40 +0200 Subject: [PATCH] Use fused gemv+add only for TG --- ggml/src/ggml-cuda.cu | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 7e2da7f8..40ac388d 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -2067,7 +2067,7 @@ static int ggml_cuda_mul_mat_q(ggml_backend_cuda_context & ctx, const ggml_tenso auto stream = ctx.stream(); - auto fusion = ctx.fusion; + auto fusion = ctx.fusion && src1->ne[1] == 1; auto ne10_padded = GGML_PAD(src1->ne[0], MATRIX_ROW_PADDING); auto nb10_padded = ne10_padded*sizeof(block_q8_1)/QK8_1;