[BugFix] fix generating endlessly bug when reasoner grammar & spec decoding are enabled

[justadogistaken]

Motivation

Related to #9187 and #11550.

When reasoner grammar and spec decoding are enabled at the same time.
The model will generate <think> first, and nothing would be generated till reach max_tokens.

I found that the problem comes from allocate_token_bitmask.fill_(0) and reasoner grammar.

def traverse_tree(
    retrieve_next_token: torch.Tensor,
    retrieve_next_sibling: torch.Tensor,
    draft_tokens: torch.Tensor,
    grammar: BaseGrammarObject,
    allocate_token_bitmask: torch.Tensor,
):
    """
    Traverse the tree constructed by the draft model to generate the logits mask.
    """
    assert (
        retrieve_next_token.shape == retrieve_next_sibling.shape == draft_tokens.shape
    )

    allocate_token_bitmask.fill_(0)

    def dfs(
        curr: int,
        retrieve_next_token: torch.Tensor,
        retrieve_next_sibling: torch.Tensor,
        parent_pos: int,
    ):

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[gemini-code-assist]

Summary of Changes

Hello @justadogistaken, 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 resolves a significant bug that caused models to enter an infinite loop of empty string generation when using reasoner grammar alongside speculative decoding. The fix involves refining how grammar states are managed and copied, particularly during vocabulary mask generation, and introducing more precise control over when tokens are masked. These changes ensure that the grammar's state is correctly preserved and that masking logic is applied only when appropriate, leading to more stable and predictable model output.

Highlights

• Bug Fix: Addresses a critical bug where enabling both reasoner grammar and speculative decoding could lead to the model generating an endless sequence of empty strings ('') until max_tokens is reached. This was traced to incorrect vocabulary masking logic during the reasoning phase.
• Grammar State Management: Introduced a new abstract method accept_token_length to BaseGrammarObject and implemented it in XGrammarGrammar and ReasonerGrammarObject to accurately track the number of accepted tokens, improving grammar state handling.
• Precise Vocabulary Masking: Modified ReasonerGrammarObject to correctly manage vocabulary masking based on whether the model is currently in a reasoning phase (is_in_reasoning) or has just generated the </think> end token. This prevents unintended masking during critical reasoning steps.
• Preventing Side Effects: Ensured that the grammar object is copied (req.grammar.copy()) before being used in the dfs function within generate_token_bitmask. This prevents modifications to the grammar state during bitmask generation from affecting other parts of the system, such as speculative verification.
• Robust Rollback Logic: Added a condition (grammar.accept_token_length() > 0) to the dfs function's rollback logic to prevent attempting to rollback when no tokens have been accepted by the grammar, enhancing 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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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 effectively addresses a critical bug causing an infinite loop when reasoner grammar and speculative decoding are enabled simultaneously. The core of the fix involves introducing accept_token_length() to prevent erroneous rollbacks on the grammar, and ensuring that a copy of the grammar object is used during speculative tree traversal to preserve its state. The changes to ReasonerGrammarObject to correctly manage its state during the reasoning phase are also well-implemented. Overall, the logic is sound and the fix appears robust. I have a couple of minor suggestions to improve code clarity and consistency.

[justadogistaken]

Test

Problem Solved with this pr. And this has been deployed in production.

Request

curl 'http://localhost:30000/v1/chat/completions' \
-H 'Content-Type: application/json' \
-d '{
    "model": "Qwen/Qwen3-8B", 
    "messages": [ 
        {
            "role": "system",
            "content": "You are a highly advanced and versatile AI assistant."
        },
        {
            "role": "user",
            "content": "Tell me about the latest advancements in AI."
        }
    ],
    "max_tokens": 1000,
    "response_format": {
        "type": "json_object"
    }
}'

[Leoyzen]

Will this PR being merge?I've use this pr for quite a long time to prevent the hang issues.

[justadogistaken]

@hnyls2002 @merrymercy @Ying1123 @kssteven418 Hi, could you help review this pr?

[LorrinWWW]

@justadogistaken could you resolve the conflicts, tysm!

[justadogistaken]

@LorrinWWW The problem seems to be solved in the latest version. I will close the pr. But I found there is another problem. Could you help review this?

ref #14480