MindSpore Serving Installation

View Source On Gitee

Installation

Currently, MindSpore Serving can be deployed only in the Linux environment.

MindSpore Serving wheel packages are common to various hardware platforms(Nvidia GPU, Ascend 910/310P/310, CPU). The inference task depends on the MindSpore or MindSpore Lite inference framework. We need to select one of them as the Serving Inference backend. When these two inference backend both exist, Mindspore Lite inference framework will be used.

MindSpore and MindSpore Lite have different build packages for different hardware platforms. The following table lists the target devices and model formats supported by each build package.

Inference backend

Build platform

Target device

Supported model formats

MindSpore

Nvidia GPU

Nvidia GPU

MindIR

Ascend

Ascend 910

MindIR

MindSpore Lite

Nvidia GPU

Nvidia GPU, CPU

MindIR_Lite

Ascend

Ascend 310P/310, CPU

MindIR_Lite

CPU

CPU

MindIR_Lite

When MindSpore is used as the inference backend, MindSpore Serving supports the Ascend 910 and Nvidia GPU environments. Ascend 910 and GPU environment only supports the MindIR model format.

Due to the dependency between MindSpore Serving and MindSpore, please follow the table below, download and install the corresponding MindSpore verision from MindSpore download page.

MindSpore Serving Version

Branch

MindSpore Version

2.0.0

r2.0

2.0.0

1.9.0

r1.9

1.9.0

1.8.0

r1.8

1.8.0, 1.8.1

1.7.0

r1.7

1.7.0

For details about how to install and configure MindSpore, see Installing MindSpore and Configuring MindSpore.

When MindSpore Lite is used as the inference backend, MindSpore Serving supports Ascend 310P/310, Nvidia GPU and CPU environments. Only the MindIR_Lite model formats is supported. Models of format MindIR exported from MindSpore or models of other frameworks need be be converted to MindIR_Lite format by the MindSpore Lite conversion tool. The MindIR_Lite models converted from Ascend310 and Ascend310P environments are different, and the MindIR_Lite models must be running on the corresponding Ascend310 or Ascend310P environments. MindIR_Lite models converted from Nvidia GPU and CPU environments can be running only in the Nvidia GPU and CPU environments.

Inference backend

Running environment of Lite conversion tool

Target device of MindIR_Lite models

MindSpore Lite

Nvidia GPU, CPU

Nvidia GPU, CPU

Ascend 310

Ascend 310

Ascend 310P

Ascend 310P

For details about how to compile and install MindSpore Lite, see the MindSpore Lite Documentation. We should configure the environment variable LD_LIBRARY_PATH to indicates the installation path of libmindspore-lite.so.

We can install MindSpore Serving either by pip or by source code.

Installation by pip

Perform the following steps to install Serving:

If use the pip command, download the .whl package from the MindSpore Serving page and install it.

pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/{version}/Serving/{arch}/mindspore_serving-{version}-{python_version}-linux_{arch}.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple
  • {version} denotes the version of MindSpore Serving. For example, when you are downloading MindSpore Serving 1.1.0, {version} should be 1.1.0.

  • {arch} denotes the system architecture. For example, the Linux system you are using is x86 architecture 64-bit, {arch} should be x86_64. If the system is ARM architecture 64-bit, then it should be aarch64.

  • {python_version} spcecifies the python version for which MindSpore Serving is built. If you wish to use Python3.7, {python_version} should be cp37-cp37m. If Python3.8 is used, it should be cp38-cp38. If Python3.9 is used, it should be cp39-cp39. Please use the same Python environment whereby MindSpore Serving is installed.

Installation by Source Code

Install Serving using the source code.

git clone https://gitee.com/mindspore/serving.git -b master
cd serving
bash build.sh

For the bash build.sh above, we can add -jn, for example -j16, to accelerate compilation. By adding -S on option, third-party dependencies can be downloaded from gitee instead of github.

MindSpore Serving compliation depends on the MindSpore inference header files. During the preceding compilation, the MindSpore source code will be downloaded. If the MindSpore whl or MindSpore Lite package has been installed, we can run the following compilation command to avoid downloading the MindSpore source code.

git clone https://gitee.com/mindspore/serving.git -b master
cd serving
bash build.sh -p ${mindspore_path}/lib

Use the -p parameter to specify the path of MindSpore whl or MindSpore Lite package. The ${mindspore_path} indicates the installation path of MindSpore whl package or runtime path in MindSpore Lite tar package.

After the build is complete, find the .whl installation package of Serving in the serving/build/package/ directory and install it.

pip install mindspore_serving-{version}-{python_version}-linux_{arch}.whl

Installation Verification

Run the following commands to verify the installation. Import the Python module. If no error is reported, the installation is successful.

from mindspore_serving import server