diff --git a/ggml-cuda.cu b/ggml-cuda.cu index 15729691a..dc3237a45 100644 --- a/ggml-cuda.cu +++ b/ggml-cuda.cu @@ -522,7 +522,7 @@ static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUA #define MUL_MAT_SRC1_COL_STRIDE 128 -#define MAX_STREAMS 8 +#define MAX_STREAMS 4 static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } }; struct ggml_tensor_extra_gpu {