msModelSlim Installation Guide

Installation Description

This document describes how to install msModelSlim. Currently, three installation methods are supported: installation from PyPI, installation through a .whl package, and building from source.

Preparations

Prepare the Python environment: Python 3.8 or later is required.

Building from Source

[!NOTE]Note

When using the msModelSlim command line tool, do not run commands in the msModelSlim source code directory. Doing so may cause conflicts between the source code path and the installation path when Python imports modules, leading to command execution failures.
If you encounter errors when installing msmodelslim, see the FAQ for solutions. If the issue persists, submit an issue with your operating environment and complete error logs attached. We will troubleshoot the issue for you as soon as possible.

Installation on Atlas A2 Training and Inference Products and Atlas A3 Training and Inference Products

# 1. Clone the msModelSlim repository using git clone.
git clone https://gitcode.com/Ascend/msmodelslim.git

# 2. Go to the msModelSlim directory and run the installation script.
cd msmodelslim
bash install.sh

Installation on Atlas 300I Duo Products

Prerequisites: CANN has been installed and environment variables have been set.

Notes:

The Atlas 300I Duo card supports quantization only with a single processor on a single device.
To perform sparse quantization and compression, install CANN 8.2.RC1 or later.

Download the CANN package. Select the AArch64 or x86_64 version based on your system architecture. For details about the installation method, see the CANN Installation Guide.

# 1. Clone the msModelSlim repository using git clone.
git clone https://gitcode.com/Ascend/msmodelslim.git

# 2. Go to the msModelSlim directory and run the installation script.
cd msmodelslim
bash install.sh

# Note: To perform sparse quantization and compression, proceed with the following operations.
# 3. Go to the site-packages package management directory under the Python environment, where ${python_envs} specifies the Python environment path.
cd ${python_envs}/site-packages/msmodelslim/pytorch/weight_compression/compress_graph/  
# In the following example, /usr/local/ is the user directory and the Python version is 3.11.10.
cd /usr/local/lib/python3.11/site-packages/msmodelslim/pytorch/weight_compression/compress_graph/

# 4. Build the weight_compression component, where ${install_path} specifies the installation directory of the CANN software.
sudo bash build.sh ${install_path}/ascend-toolkit/latest

# 5. The build operation in the previous step generates the build directory. Grant relevant permissions to the build directory.
chmod -R 550 build

Installation from `PyPI`

pip install msmodelslim

Installation Through a `.whl` Package

Download the msModelSlim .whl software package by referring to section "Wheel Package Downloads" in Release Notes.

After obtaining the .whl package, run the following command to install it:

sha256sum {name}.whl # Verify the .whl package. If the checksums match, the .whl package is not damaged during download.

pip install ./msmodelslim-{version}-py3-none-any.whl # Install the .whl package.

Post-installation Configuration

For Ascend NPU devices, perform the following operations.

Installing CANN