[Bug] 910b multi-card reasoning is very slow.

[the-nine-nation]

Checklist

• 1. I have searched related issues but cannot get the expected help.
• 2. The bug has not been fixed in the latest version.
• 3. Please note that if the bug-related issue you submitted lacks corresponding environment info and a minimal reproducible demo, it will be challenging for us to reproduce and resolve the issue, reducing the likelihood of receiving feedback.

Describe the bug

我用双卡910b进行推理时，速度比单卡910b慢了大约30%
用的torch_npu，测试模型是qwen2.5 7b instruct
单条长回复，单卡大约32token/s，双卡只有22token/s

Reproduction

我运行服务命令：lmdeploy serve api_server --backend pytorch --device ascend /home/ma-user/work/qwen2-7b --server-port 6007 --tp 2 --cache-max-entry-count=0.9

Environment

Warning : ASCEND_HOME_PATH environment variable is not set.
/home/ma-user/anaconda3/envs/py39/lib/python3.9/site-packages/torch_npu/utils/path_manager.py:82: UserWarning: Warning: The /home/ma-user/work owner does not match the current user.
  warnings.warn(f"Warning: The {path} owner does not match the current user.")
[W compiler_depend.ts:623] Warning: expandable_segments currently defaults to false. You can enable this feature by `export PYTORCH_NPU_ALLOC_CONF = expandable_segments:True`. (function operator())
[W compiler_depend.ts:631] Warning: expandable_segments feature is not supportted                     and the possible cause is that driver and firmware packages do not match. (function operator())
sys.platform: linux
Python: 3.9.20 | packaged by conda-forge | (main, Sep 22 2024, 14:02:18) [GCC 13.3.0]
CUDA available: False
MUSA available: False
numpy_random_seed: 2147483648
GCC: gcc (GCC) 7.3.0
PyTorch: 2.1.0
PyTorch compiling details: PyTorch built with:
  - GCC 10.2
  - C++ Version: 201703
  - Intel(R) MKL-DNN v3.1.1 (Git Hash 64f6bcbcbab628e96f33a62c3e975f8535a7bde4)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - LAPACK is enabled (usually provided by MKL)
  - NNPACK is enabled
  - CPU capability usage: NO AVX
  - Build settings: BLAS_INFO=open, BUILD_TYPE=Release, CXX_COMPILER=/opt/rh/devtoolset-10/root/usr/bin/c++, CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 -fabi-version=11 -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOCUPTI -DLIBKINETO_NOROCTRACER -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=old-style-cast -Wno-invalid-partial-specialization -Wno-unused-private-field -Wno-aligned-allocation-unavailable -Wno-missing-braces -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=open, TORCH_DISABLE_GPU_ASSERTS=ON, TORCH_VERSION=2.1.0, USE_CUDA=OFF, USE_CUDNN=OFF, USE_EIGEN_FOR_BLAS=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=OFF, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=OFF, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF, 

TorchVision: 0.16.0
LMDeploy: 0.6.0+
transformers: 4.44.2
gradio: 4.44.0
fastapi: 0.115.0
pydantic: 2.9.2
triton: Not Found

Error traceback

No response

[jinminxi104]

We think this issue can be solved by supporting MatmulAllReduce op.
We'll support MatmulAllReduce in Oct.