EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Bugfix] fix custom op GmmSwigluQuantWeightNzTensorList (#4593)

Browse Source 
### What this PR does / why we need it?

1. Fixes the environment path used to locate custom op shared libraries.
2. Uses empty tensor initialization for op outputs instead of
zero-initialization for better efficiency.



- vLLM version: v0.11.2
- vLLM main: https://github.com/vllm-project/vllm/commit/v0.11.2

---------

Signed-off-by: QianChenxi <chenxi.qian.cq@outlook.com>
This commit is contained in:
Chenxi Qian
2025-12-02 22:02:04 +08:00
committed by GitHub
parent b84c9afbf5
commit 4588cdac02
6 changed files with 18 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
csrc/CMakeLists.txt
View File

@@ -15,7 +15,7 @@ option(BUILD_OPEN_PROJECT "Build open ascend ops project." ON)
option(ENABLE_CCACHE "Enable ccache capability" ON)
set(ASCEND_COMPUTE_UNIT "ascend910b" CACHE STRING "soc that need to be compiled")
set(ASCEND_OP_NAME "ALL" CACHE STRING "operators that need to be compiled")
set(VENDOR_NAME "customize" CACHE STRING "vendor name")
set(VENDOR_NAME "vllm-ascend" CACHE STRING "vendor name")
include(cmake/config.cmake)
include(cmake/func.cmake)

1
csrc/build_aclnn.sh
View File

@@ -31,4 +31,3 @@ bash build.sh -n $CUSTOM_OPS -c $SOC_ARG
# install custom ops to vllm_ascend/_cann_ops_custom
./output/CANN-custom_ops*.run --install-path=$ROOT_DIR/vllm_ascend/_cann_ops_custom
source $ROOT_DIR/vllm_ascend/_cann_ops_custom/vendors/customize/bin/set_env.bash

6
csrc/torch_binding.cpp
View File

@@ -568,9 +568,9 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor> grouped_matmul_swiglu_quant_weigh
int m = x_size[0];
int k = x_size[1];
at::Tensor output = at::zeros({m, n/2}, x.options().dtype(at::kChar));
at::Tensor output_scale = at::zeros({m}, x.options().dtype(at::kFloat));
at::Tensor output_offset = at::zeros({m}, x.options().dtype(at::kFloat));
at::Tensor output = at::empty({m, n/2}, x.options().dtype(at::kChar));
at::Tensor output_scale = at::empty({m}, x.options().dtype(at::kFloat));
at::Tensor output_offset = at::empty({m}, x.options().dtype(at::kFloat));
EXEC_NPU_CMD(
aclnnGroupedMatmulSwigluQuantWeightNzTensorList,