From 39ddda27f43e487a4c05375003f5bc7acb38a742 Mon Sep 17 00:00:00 2001 From: slaren Date: Sat, 30 Sep 2023 17:17:38 +0200 Subject: [PATCH] disable fp16 mat mul completely with multi GPU --- ggml-cuda.cu | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 2fd15957f..989c419cd 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -6118,7 +6118,7 @@ inline void ggml_cuda_op_mul_mat_cublas( const int compute_capability = g_compute_capabilities[id]; - if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && ldc == row_diff) { + if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) { // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32 half * src0_as_f16 = nullptr; size_t src0_as = 0;