From c3b75c531c5d5e295d5c3ba846bb37def72df3a7 Mon Sep 17 00:00:00 2001 From: Kawrakow Date: Wed, 19 Mar 2025 13:03:59 +0100 Subject: Prevent FlashMLA-1 from running on CUDA (#268) as it is not supported. Co-authored-by: Iwan Kawrakow --- ggml/src/ggml-cuda.cu | 3 +++ 1 file changed, 3 insertions(+) (limited to 'ggml/src') diff --git a/ggml/src/ggml-cuda.cu b/ggml/src/ggml-cuda.cu index 453bde0c..e488c9bc 100644 --- a/ggml/src/ggml-cuda.cu +++ b/ggml/src/ggml-cuda.cu @@ -3579,6 +3579,9 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons return (op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) || (op->src[1]->type == GGML_TYPE_Q8_0 && op->src[2]->type == GGML_TYPE_Q8_0); } + if (op->src[1]->ne[0] > 256) { + return false; + } if (op->src[0]->ne[0] == 64 && op->src[1]->type == GGML_TYPE_F16) { return true; } -- cgit v1.2.3