[BugFix] Work-around incremental detokenization edge case error #19449
Conversation
Signed-off-by: Nick Hill <nhill@redhat.com>
njhill
requested review from
WoosukKwon,
alexm-redhat,
comaniac,
robertgshaw2-redhat and
ywang96
as code owners
|
👋 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 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 🚀 |
![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 @njhill, 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 provides a work-around for a specific detokenization bug that can occur with the fast detokenizer when processing certain token sequences that cause the tokenizers library's DecodeStream to fail. The fix involves catching the specific error and resetting the stream, preventing the failure. A new test case is included to demonstrate and verify the fix.
Highlights
- Bug Fix: Addresses an edge case in the
FastIncrementalDetokenizerwhere the underlyingtokenizerslibrary'sDecodeStreamcan encounter an 'Invalid prefix encountered' error due to non-monotonic or invalid UTF-8 output from certain token sequences. This fixes issue #17448. - Error Handling: Implements a protected step method (
_protected_step) inFastIncrementalDetokenizerthat wraps theDecodeStream.stepcall in a try-except block. If the specific 'Invalid prefix encountered' exception is caught, the decode stream is reset to recover. - New Test Case: Adds a new test file (
tests/v1/engine/test_fast_incdec_prefix_err.py) with a specific sequence of tokens that reproduces the detokenization error using thegoogle/gemma-3-1b-ittokenizer, ensuring the fix is effective.
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 configureGemini 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 PR introduces a workaround for an edge case error in the tokenizers library during incremental detokenization. The fix involves catching the "Invalid prefix encountered" error, resetting the DecodeStream, and attempting to continue. A new test case validates this recovery mechanism. The code changes are well-localized and clear.
Signed-off-by: Nick Hill <nhill@redhat.com>
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
Signed-off-by: Nick Hill <nhill@redhat.com>
Signed-off-by: Nick Hill <nhill@redhat.com>
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
…-project#19449) Signed-off-by: Nick Hill <nhill@redhat.com>
…-project#19449) (#1484) Cherry-pick of vllm-project#19449 which resolves the issue with tokenizers package --------- Signed-off-by: Nick Hill <nhill@redhat.com> Signed-off-by: Karol Damaszke <kdamaszke@habana.ai> Co-authored-by: Nick Hill <nhill@redhat.com>
…-project#19449) Signed-off-by: Nick Hill <nhill@redhat.com>
…-project#19449) Signed-off-by: Nick Hill <nhill@redhat.com>
…-project#19449) Signed-off-by: Nick Hill <nhill@redhat.com>
…-project#19449) (vllm-project#243) Signed-off-by: Nick Hill <nhill@redhat.com> Co-authored-by: Nick Hill <nhill@redhat.com>
2 participants
I will be opening an issue for this in the tokenizers repo too.
Thanks very much to @fpaupier for providing a clear/detailed reproducer for the issue.
Fixes #17448.