diff options
| author | Johannes Gäßler <johannesg@5d6.de> | 2023-09-11 22:58:41 +0200 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2023-09-11 22:58:41 +0200 |
| commit | 89e89599fd095172f8d67903b5e227467420f036 (patch) | |
| tree | 8628a31b9b91b0affdc1e5ff60eab7fe5c610497 | |
| parent | d54a4027a6ebda98ab0fef7fa0c2247d0bef132a (diff) | |
CUDA: fix mul_mat_q not used for output tensor (#3127)
| -rw-r--r-- | ggml-cuda.cu | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 9e9eac48..a14e2362 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -6291,7 +6291,7 @@ void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_ ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false); } } else { - if (src1->backend == GGML_BACKEND_GPU && g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) { + if (g_mul_mat_q && ggml_is_quantized(src0->type) && min_compute_capability >= MIN_CC_DP4A) { ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true); } else { ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false); |