[bugfix] fix hf token fetch and deletion logic#122
Conversation
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @slin1237, 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 delivers a series of crucial bug fixes and reliability enhancements for the model-agent, primarily focusing on improving HuggingFace model download stability, authentication, and lifecycle management. The changes aim to prevent common issues like rate limiting, authentication failures, and stuck downloads, making the agent more robust for production environments.
Highlights
- HuggingFace Token Authentication Fix: I've addressed a critical bug where HuggingFace tokens were retrieved but not properly passed to download operations, resolving HTTP 401 errors for gated models. This is now handled by a new
WithDownloadTokenoption. - Enhanced Rate Limiting Handling: I've implemented robust rate limit handling for HuggingFace API calls, including exponential backoff with jitter, parsing of
Retry-Afterheaders, and increased retry attempts. A staggered start mechanism based on node names has also been added to prevent 'thundering herd' issues in large clusters. - Download Cancellation Support: I've added support for cancelling in-progress downloads. The model agent now tracks active downloads and can gracefully cancel them when a model resource is deleted, preventing stuck downloads and improving resource management.
- Flexible Secret Key Configuration: The secret key name used to fetch HuggingFace tokens is now configurable via model parameters (defaults to 'token'), allowing for greater flexibility in secret management. Documentation has been updated to reflect this.
- Model Deletion Path Consistency: I've fixed an issue where model deletion used an inconsistent path construction, ensuring that
getDestPath()is used uniformly for both download and delete operations, leading to proper cleanup. - General Stability Improvements: I've included fixes for nil pointer dereferences in progress logging and added checks for incomplete files before renaming during downloads, enhancing overall stability.
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
-
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. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces several important bug fixes and improvements for the model-agent, primarily focused on enhancing reliability in production environments. Key changes include fixing Hugging Face token authentication, adding robust rate-limiting with exponential backoff, and implementing download cancellation. The changes are well-structured and address critical issues.
My review has identified a few areas for improvement:
- A potential bug in the startup jitter calculation that could cause it to not work for some nodes.
- A race condition in the download cancellation logic due to the use of
time.Sleepfor synchronization. - A misleading comment in the download logic.
Addressing these points will further improve the robustness and maintainability of the agent.
| // Wait a bit for download to stop | ||
| time.Sleep(2 * time.Second) |
There was a problem hiding this comment.
Using a hardcoded time.Sleep is not a reliable way to ensure a download has been cancelled and has stopped writing to files. This creates a race condition where file deletion might start while the download is still in progress, potentially causing errors or leaving the filesystem in an inconsistent state.
A more robust approach would involve a proper synchronization mechanism, such as a sync.WaitGroup, to confirm that the download goroutine has fully terminated before proceeding with deletion.
| hash := 0 | ||
| for _, c := range cfg.nodeName { | ||
| hash = hash*31 + int(c) | ||
| } |
There was a problem hiding this comment.
The current hash calculation for the jitter delay can result in a negative hash value due to integer overflow. In Go, the % operator with a negative number yields a negative result (e.g., -5 % 30 is -5). Consequently, time.Sleep will receive a negative duration and return immediately, defeating the purpose of the jitter for some node names.
To ensure the delay is always non-negative, the hash should be calculated using an unsigned integer type.
| hash := 0 | |
| for _, c := range cfg.nodeName { | |
| hash = hash*31 + int(c) | |
| } | |
| var hash uint32 | |
| for _, c := range cfg.nodeName { | |
| hash = hash*31 + uint32(c) | |
| } |
|
|
||
| // Move the file to final destination | ||
| if err := os.Rename(incompletePath, destPath); err != nil { | ||
| // Check if it's a cross-device error and try copy instead |
There was a problem hiding this comment.
This comment is misleading. The code below it checks for os.IsNotExist and returns an error, but it does not handle cross-device move errors (which would be syscall.EXDEV on Unix) or attempt to copy the file as the comment suggests. The comment should be updated to accurately describe the code's behavior.
| // Check if it's a cross-device error and try copy instead | |
| // Add more context if the rename failed because the source file disappeared. |
1 participant
What type of PR is this?
/kind bug
What this PR does / why we need it:
This PR fixes multiple critical issues in the model-agent that were preventing proper model downloads and management in production environments:
1. RBAC Permissions Issue
2. HTTP 429 Rate Limiting
3. Nil Pointer Dereference
4. Model Deletion Path Mismatch
getDestPath()for both download and delete operations5. Flexible Secret Key Configuration
6. Download Cancellation Support
7. HuggingFace Token Authentication
WithDownloadTokenoption to pass tokens per downloadWhich issue(s) this PR fixes:
Fixes #
Special notes for your reviewer:
Does this PR introduce a user-facing change?