[tech debt] Revisit lora request model checker

[kouroshHakha]

Very minor PR on the LoRA request handling in OpenAIModels:

1. Make model checker O(1): This would make it easier for external integrations (e.g. Ray serve llm / dynamo) to query if model x exists in OpenAIModels data structure by just doing if model in oai_models.lora_requests.

2. Make the add and removal atomic by reusing the async lock mechanism that exist already. Currently add and remove were not atomic and could lead to race conditions if multiple concurrent request touch the same lora model for load / deletion.

[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 @kouroshHakha, 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 addresses technical debt by significantly improving the performance of LoRA (Low-Rank Adaptation) model lookups within the serving engine. By transitioning the internal storage of LoRA requests from a list to a dictionary, the system now benefits from constant-time (O(1)) access for model validation and retrieval, leading to more efficient handling of LoRA-related operations.

Highlights

• Performance Improvement: The core lora_requests collection, previously a list, has been refactored into a dictionary, enabling O(1) lookup times for LoRA models by their name. This significantly improves the efficiency of operations involving LoRA model identification and retrieval.
• Data Structure Adaptation: All operations interacting with the lora_requests collection, including adding, removing, checking existence, and retrieving LoRA models, have been updated to leverage the new dictionary structure, replacing linear scans with direct key lookups.

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. ↩

[gemini-code-assist]

Code Review

This pull request correctly implements the intended optimization of using a dictionary for LoRA requests, which improves lookup performance to O(1).

However, I've identified a critical race condition in the load_lora_adapter and unload_lora_adapter methods. These methods are not atomic, which can lead to inconsistent state or even server errors under concurrent access. I've provided a detailed comment on how to address this by using a locking mechanism similar to what's already used in resolve_lora.

[kouroshHakha]

/gemini review

[kouroshHakha]

@gemini-code-assist please re-review?

[gemini-code-assist]

Code Review

This pull request improves LoRA request handling by using a dictionary for lora_requests (O(1) lookup) and introducing atomic operations for loading/unloading adapters, preventing race conditions. Exception handling could be improved by catching Exception instead of BaseException. A pre-existing logic issue regarding unloading adapters using lora_int_id was also identified.

[gemini-code-assist]

Code Review

The pull request revisits the LoRA request model checker to improve efficiency and atomicity. It changes the data structure for storing LoRA requests from a list to a dictionary for O(1) model checking and uses async locks to ensure atomic add and removal operations. The changes look good and address the tech debt as described.

[kouroshHakha]

Changed from BaseException to Exception because @gemini-code-assist suggested which is fair.

[aarnphm]

LGTM. let's run the full suite to see if there are any breakage here.

[kouroshHakha]

@aarnphm please hit the merge button whenever you can. Thanks.