vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

musa: workaround for Guilty Lockup in cleaning src0

Browse source 
Signed-off-by: Xiaodong Ye <xiaodong.ye@mthreads.com>
This commit is contained in:
Xiaodong Ye 2024-10-25 16:19:33 +08:00
parent 958367bf53
commit 862b9598d9
1 changed files with 6 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ggml/src/ggml-cuda.cu
View file

@ -1484,7 +1484,12 @@ static void ggml_cuda_op_mul_mat(
const size_t nbytes_data = ggml_nbytes(src0); 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); 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); 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)); 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 src0 is on a temporary compute buffer (partial offloading) there may be some padding that needs to be cleared: