cuda : prevent using split buffers with 3d/4d matrices (llama/13919)

diff --git a/src/ggml-cuda/ggml-cuda.cu b/src/ggml-cuda/ggml-cuda.cu
index c442a64924303146040aa58016d59482d16b2363..009ed9048dad3c911829f2763a5c31c96fb1a396 100644 (file)
--- a/src/ggml-cuda/ggml-cuda.cu
+++ b/src/ggml-cuda/ggml-cuda.cu
@@ -2994,9 +2994,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
-                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
-                // this avoids some edge cases (and the performance would not be good anyways)
                 if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    if (a->ne[2] > 1 || a->ne[3] > 1) {
+                        return false;
+                    }
+                    // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                    // this avoids some edge cases (and the performance would not be good anyways)
                     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
                     int64_t row_low;
                     int64_t row_high;