From: R0CKSTAR Date: Mon, 28 Oct 2024 09:02:48 +0000 (+0800) Subject: musa: workaround for Guilty Lockup in cleaning src0 (llama/10042) X-Git-Tag: upstream/0.0.1642~257 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=5db70fabf1279d16dd360d994eee7781e71836e7;p=pkg%2Fggml%2Fsources%2Fggml musa: workaround for Guilty Lockup in cleaning src0 (llama/10042) Signed-off-by: Xiaodong Ye --- diff --git a/src/ggml-cuda.cu b/src/ggml-cuda.cu index 21c9f5e3..217df968 100644 --- a/src/ggml-cuda.cu +++ b/src/ggml-cuda.cu @@ -1484,14 +1484,19 @@ static void ggml_cuda_op_mul_mat( const size_t nbytes_data = ggml_nbytes(src0); const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING); dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), nbytes_data + nbytes_padding); + // TODO: remove this for MUSA once the Guilty Lockup issue is resolved +#ifndef GGML_USE_MUSA CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd, 0, nbytes_data + nbytes_padding, stream)); +#else // GGML_USE_MUSA + CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream)); +#endif // !GGML_USE_MUSA } // If src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared: if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) { const size_t nbytes_data = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00); const size_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING); - CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream)); + CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data, 0, nbytes_padding, stream)); } if (src1_on_device && src1_is_contiguous) {