[Bugfix] improve regex for hermes tool detection #20474
Conversation
|
👋 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 @XciD, 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 implements a bug fix to enhance the robustness of the Hermes tool parser. It addresses an issue where models might generate slightly malformed tool call tags, preventing successful parsing. The update ensures that the system can correctly extract tool calls even with these variations, improving overall reliability.
Highlights
- Improved Tool Call Parsing Robustness: The regular expression used in
hermes_tool_parser.pyfor detecting tool calls has been updated to be more forgiving. Specifically, it now correctly handles cases where a model might output multiple consecutive opening<tool_call>tags (e.g.,<tool_call><tool_call>...), which previously led to parsing errors. - Bug Fix for
JSONDecodeError: This change resolves aJSONDecodeErrorthat occurred when the Hermes parser encountered malformed tool call outputs from certain models, such as Qwen, ensuring that tool calls are successfully extracted and processed.
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 addresses a bug in the Hermes tool parser where multiple <tool_call> tags were causing JSON decoding errors. The change to the regex is a good step forward. I've identified a potential issue with the new regex where it might not handle whitespace between the repeated tags correctly, and I've suggested a more robust pattern to address this. With this change, the parser should be able to handle malformed tool call outputs more reliably.
mergify
bot
added
documentation
Signed-off-by: Adrien <adrien@huggingface.co> Co-authored-by: oOraph
| find(self.tool_call_start_token)] | ||
| tool_call_start = model_output.find(self.tool_call_start_token) | ||
| content = (model_output[:tool_call_start] | ||
| if tool_call_start >= 0 else None) |
There was a problem hiding this comment.
This condition is sort of a duplicate of the condition on line 112.
The one on line 112 is a bit clearer, can we keep that one?
There was a problem hiding this comment.
hi @hmellor :), I don't understand what you mean, can you explain ?
test = 'whatever'
lookup = 'sthnotfound'
print(test[:test.find(lookup)]) # print(test[:-1]) -> 'whateve' -> not None, so if we remove this check this will be the content value no ?
Essential Elements of an Effective PR Description Checklist
Purpose
Improve the hermes parser to be less strict. Some model like qwen can sometimes output double xml tag for tool_call.
Generated output:
Stack
Test Plan
Difficult to give you a buggy prompt as our contains sensitive information.
Test Result
Result completion with buggy prompt:
Before
{ "id": "chatcmpl-vgunWp", "object": "chat.completion", "created": 1751612418, "model": "Qwen/Qwen2.5-72B-Instruct", "choices": [ { "index": 0, "message": { "role": "assistant", "reasoning_content": null, "content": "The repository contains the following files:\n\n- `.gitattributes`\n- `.py`\n- `README.md`\n- `app.py`\n- `requirements.txt`\n\nNext, I will review the contents of these files. Let's start with the `app.py` file, which is likely the main application script.\n<tool_call>\n<tool_call>\n{\"name\": \"tool-1_hf_get_file_content\", \"arguments\": {\"file_path\": \"app.py\", \"repo_type\": \"spaces\", \"repository\": \"xcid/test\"}}\n</tool_call>", "tool_calls": [] }, "logprobs": null, "finish_reason": "stop", "stop_reason": null } ], "usage": { "prompt_tokens": 3052, "total_tokens": 3173, "completion_tokens": 121, "prompt_tokens_details": null }, "prompt_logprobs": null, "kv_transfer_params": null }After
{ "id": "chatcmpl-BGqu6p", "object": "chat.completion", "created": 1751612343, "model": "Qwen/Qwen2.5-72B-Instruct", "choices": [ { "index": 0, "message": { "role": "assistant", "reasoning_content": null, "content": "The repository contains the following files:\n\n- `.gitattributes`\n- `.py`\n- `README.md`\n- `app.py`\n- `requirments.txt`\n\nNext, I will review the contents of these files. Let's start with the `app.py` file, which is likely the main application script.\n", "tool_calls": [ { "id": "chatcmpl-tool-6b38a134cc2d4f3b84526f2936155583", "type": "function", "function": { "name": "tool-1_hf_get_file_content", "arguments": "{\"repo_type\": \"spaces\", \"repository\": \"xcid/test\", \"file_path\": \"app.py\"}" } } ] }, "logprobs": null, "finish_reason": "tool_calls", "stop_reason": null } ], "usage": { "prompt_tokens": 3052, "total_tokens": 3173, "completion_tokens": 121, "prompt_tokens_details": null }, "prompt_logprobs": null, "kv_transfer_params": null }Co-authored-by: @oOraph