fix(ci): update vllm to 0.21.0 for Dockerfile #4934
Conversation
Added wetext installation and set LD_LIBRARY_PATH.
Removed 'gptqmodel' from the requirements file.
Set environment variables for NVM and Node.js version.
Removed pre-release installation of torch, torchvision, and torchaudio, retaining only triton and torchcodec installations.
Removed installation of pre-release PyTorch packages and adjusted the installation order.
Removed installation of torchcodec from Dockerfile.
Change docker build command to use --load instead of --push for aarch64 image.
Added security options to the Docker build for ARM64.
Removed the security option from the Docker build command for aarch64.
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request updates the Dockerfile by bumping the base image version, adjusting library paths for Python 3.12, and modifying package dependencies, including the removal of Triton and specific FlashInfer components. However, several critical issues were identified: the specified versions for the base image, Transformers, and the Flash Attention wheel appear to be non-existent or incorrect, which will lead to build failures. Additionally, hardcoding a regional pip mirror is discouraged in favor of the official PyPI index to ensure global accessibility.
I am having trouble creating individual review comments. Click here to see my feedback.
xinference/deploy/docker/Dockerfile (46)
The version transformers==5.5.0 does not exist on PyPI (the current latest stable version is in the 4.x range). This will cause the Docker build to fail. Please verify the intended version.
xinference/deploy/docker/Dockerfile (52)
The wheel URL references cu130 (CUDA 13.0) and torch2.11. Neither CUDA 13.0 nor PyTorch 2.11 have been released yet. This URL is likely invalid and will cause the build to fail.
xinference/deploy/docker/Dockerfile (1)
The tag v0.21.0 for vllm/vllm-openai does not appear to be an official release tag (vLLM is currently in the v0.7.x range). Please verify if this is a typo or if it refers to a specific non-official image.
xinference/deploy/docker/Dockerfile (29)
Hardcoding the PIP_INDEX to a regional mirror is not recommended for a general-purpose Dockerfile as it may be inaccessible or slower for users outside of China. It is better to keep the official PyPI index as the default and allow users to override it via a build argument.
ARG PIP_INDEX=https://pypi.org/simple
qinxuye
changed the title
3 participants
No description provided.