Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-26 14:20:31 +01:00
Code Issues Packages Projects Releases Wiki Activity

musa: workaround for Guilty Lockup in cleaning src0 (#10042)

Browse Source 
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
This commit is contained in:
R0CKSTAR 2024-10-28 17:02:48 +08:00 committed by GitHub
parent 8125e6cbfc
commit 524afeec9d
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 6 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-cuda.cu
View File

@ -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) {