From 1ea1089b6b2e0a8c963b3d1242d3513856a4e285 Mon Sep 17 00:00:00 2001 From: Harkirat Gill Date: Tue, 6 Oct 2026 15:04:05 -0400 Subject: [PATCH] ggml-cuda: use per-thread stream for buffer-init padding memset (llama/28782) * ggml-cuda: use per-thread stream for buffer-init padding memset * ci : re-enable test-backend-ops -j for ROCm --- ggml/src/ggml-cuda/ggml-cuda.cu | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu index 67b5e16d8..0f88dfe88 100644 --- a/ggml/src/ggml-cuda/ggml-cuda.cu +++ b/ggml/src/ggml-cuda/ggml-cuda.cu @@ -762,7 +762,8 @@ static enum ggml_status ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer if (padded_size > original_size) { ggml_cuda_set_device(ctx->device); - CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size)); + CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, cudaStreamPerThread)); + CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread)); } } return GGML_STATUS_SUCCESS;