[WIP] Draft to remove torch.cuda calls and use Platform APIs.

[jikunshang]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

Test Plan

Test Result

(Optional) Documentation Update

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @jikunshang, 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 initiates a significant refactoring effort to abstract direct torch.cuda calls behind a new Platform API. The primary goal is to enhance multi-platform compatibility, particularly for XPU devices, by centralizing device-specific operations. This change makes the codebase more modular and easier to extend to new hardware backends.

Highlights

• Platform Abstraction: Introduced new abstract methods (empty_cache, reset_peak_memory_stats, mem_get_info, memory_stats, memory_reserved, synchronize) to the Platform interface (vllm/platforms/interface.py) to standardize device-specific operations across different hardware backends.
• CUDA and XPU Implementations: Provided concrete implementations for these new platform methods in vllm/platforms/cuda.py (wrapping torch.cuda calls) and vllm/platforms/xpu.py (wrapping torch.xpu calls), enabling consistent interaction with device functionalities.
• Migration to Platform API: Replaced numerous direct torch.cuda calls with calls to the current_platform API across vllm/utils/__init__.py, vllm/v1/worker/gpu_model_runner.py, and vllm/v1/worker/gpu_worker.py for memory management, device synchronization, and device property retrieval, centralizing device interactions.
• XPU Worker Unification: Updated vllm/platforms/xpu.py to use the generic vllm.v1.worker.gpu_worker.Worker class, streamlining the worker implementation for XPU devices and reducing code duplication.
• Platform-Agnostic Device Initialization: Modified vllm/v1/worker/gpu_worker.py to use current_platform.device_name for creating torch.device instances, further decoupling device initialization from specific CUDA calls and improving multi-platform compatibility.

Using Gemini Code Assist

The 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 in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

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 .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and 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 to provide feedback.

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

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @jikunshang.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[gemini-code-assist]

Code Review

This pull request effectively refactors device-specific torch.cuda calls to use a new Platform API, which is a great step towards supporting multiple hardware backends. The changes are consistent and well-abstracted in most places.

I've identified one high-severity issue in vllm/v1/worker/gpu_worker.py where platform-specific setup logic is hardcoded for CUDA and XPU, which could be incorrect for XPU and is not maintainable. I've suggested moving this logic into the respective Platform classes for better abstraction and correctness.

[gemini-code-assist]

This condition hardcodes device types 'cuda' and 'xpu', which is not easily extensible to other platforms like ROCm. The environment variables being set within this block (TORCH_NCCL_AVOID_RECORD_STREAMS and NCCL_ASYNC_ERROR_HANDLING) are specific to NCCL, which is used by the CUDA backend. It's not clear if these are applicable or correct for the XPU backend, which uses ccl or xccl. Applying NCCL-specific workarounds to other platforms could lead to unexpected behavior or bugs.

A better approach would be to abstract this platform-specific setup into the Platform classes. For example, you could add an on_worker_init() method to the Platform interface and implement it in CudaPlatform and XpuPlatform with their respective environment variable settings. This would make the code more modular, maintainable, and less prone to errors when adding new hardware backends.