Bumps vllm version to 0.13.0 and Dockerfile to CUDA 12.9.
#1372
Conversation
Summary of ChangesHello @finbarrtimbers, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed! This pull request performs a significant dependency upgrade, moving the Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. Using Gemini Code AssistThe full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips. Invoking Gemini You can request assistance from Gemini at any point by creating a comment using either
Customization To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a Limitations & Feedback Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here. You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension. Footnotes
|
![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 vllm dependency to version 0.13.0 and configures torch to use CUDA 13.0 wheels. The changes are correctly reflected in pyproject.toml, and the lock files requirements.txt and uv.lock have been updated accordingly. The new transitive dependencies introduced by vllm and the updated nvidia-* packages for CUDA 13.0 compatibility are all present in the lock files. The changes appear correct and consistent for a dependency upgrade. I have no specific code change requests, but I recommend adding a link to the vllm v0.13.0 release notes in the pull request description for better context for future readers.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 4b2d76e820
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
pyproject.toml
Outdated
| { index = "pytorch-cu130", marker = "platform_system != 'Darwin'"}, | ||
| ] | ||
|
|
||
| [[tool.uv.index]] | ||
| name = "pytorch-cu129" | ||
| url = "https://download.pytorch.org/whl/cu129" | ||
| name = "pytorch-cu130" |
There was a problem hiding this comment.
Switching the torch index to cu130 means torch.version.cuda will report 13.0, but the Docker image still builds from CUDA 12.8 (Dockerfile:1 uses nvidia/cuda:12.8.0-devel-ubuntu22.04). Any CUDA extension that compiles inside the image (e.g., deepspeed ops or other torch.utils.cpp_extension builds) will fail the CUDA version check because nvcc is 12.x while torch is 13.x. This breaks training/experiments that compile CUDA ops in the container unless the base image is upgraded to CUDA 13 or torch is kept on cu12.
Useful? React with 👍 / 👎.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This fixes a circular import error in torchvision when running tests on CPU-only CI environments. The issue was that torchvision was being pulled from PyPI instead of the PyTorch cu130 index, causing a version mismatch with torch. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
CUDA 13.0 requires newer drivers than available on Beaker clusters. Reverting to CUDA 12.9 (cu129) which is compatible with the existing infrastructure. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Required to run this script from macOS where vllm is not available. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The tool arguments were changed in #1340 but this script wasn't updated. Now uses the new format: - --tools python (instead of --tools code search) - --tool_call_names code - --tool_configs '{"api_endpoint": "...", "timeout": 3}' - --tool_parser_type legacy Note: Search tool is not yet implemented in the new tools system. Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
finbarrtimbers
changed the title
vllm version to 0.13.0 and torch to CUDA 13.0. vllm version to 0.13.0 and Dockerfile to CUDA 12.9.
| "tensorboard>=2.18.0", | ||
| "torch>=2.9.0,<2.10", | ||
| "torchvision>=0.24.0; platform_system != 'Darwin'", | ||
| "torchaudio>=2.9.0; platform_system != 'Darwin'", |
There was a problem hiding this comment.
Why do we need these explicitly now?
Ran single GPU GRPO (Beaker), single GPU GRPO with tools (Beaker), multi-node GRPO (Beaker), and DPO (Beaker) scripts. All succeeded.