Add Cache For XGrammar Backend#933
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 @flyinglandlord, 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 significantly optimizes the XGrammarBackend by implementing a caching mechanism for compiled grammars and JSON schemas. It also introduces a new Triton-based kernel for highly efficient, batch-level application of token bitmasks, leading to improved performance for grammar-guided generation.
Highlights
- Caching XGrammar Grammars: Introduced a caching mechanism within
XgrammarBackendto store and reuse compiledxgrammargrammars and JSON schemas, reducing redundant compilation overhead for repeated requests. - XGrammarGrammar Wrapper: Created a new
XGrammarGrammarclass (lines 33-129) to encapsulatexgrammar.GrammarMatcherlogic and state, providing a consistent interface for managing grammar state, token acceptance, and facilitating the new caching strategy. - Optimized Vocab Masking: Transitioned from per-request bitmask allocation to a batch-level
vocab_mask(lines 146, 279-282), which is now allocated once per batch and filled for each request, enabling more efficient processing. - Triton-Accelerated Bitmask Application: Added a new Triton kernel (
apply_token_bitmask_inplace_tritoninbit_mask_ops.py) to efficiently apply the batch-level token bitmask to logits on CUDA devices (lines 187, 233), significantly speeding up the masking process.
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 a caching mechanism for compiled XGrammar objects and a new XGrammarGrammar class. Key areas for review include a bug in the JSON schema cache key, the handling of CPU-side vocabulary mask application, and opportunities for code deduplication and adding docstrings.
lightllm/server/router/model_infer/mode_backend/chunked_prefill/impl_for_xgrammar_mode.py
Outdated
Show resolved
Hide resolved
lightllm/server/router/model_infer/mode_backend/chunked_prefill/impl_for_xgrammar_mode.py
Show resolved
Hide resolved
| if not all_has_no_constraint: | ||
| for i, run_obj in enumerate(run_reqs): | ||
| if ( | ||
| run_obj.sampling_param.guided_grammar is not None | ||
| or run_obj.sampling_param.guided_json is not None | ||
| ): | ||
| if first_grammar is None: | ||
| first_grammar = run_obj.sampling_param.guided_grammar or run_obj.sampling_param.guided_json | ||
| self._mask_req_out_token(i, run_obj, logits[i]) |
There was a problem hiding this comment.
The logic for finding first_grammar and iterating through run_reqs is duplicated in the prefill block (lines 219-227). Consider refactoring this into a helper method to reduce redundancy.
| sample_params.xgrammar_matcher.fill_vocab_mask(self.vocab_mask, i) | ||
| return | ||
|
|
||
| def _init_req_xgrammer_matcher_infos(self, run_reqs: List[InferReq]): |
There was a problem hiding this comment.
Typo: _init_req_xgrammer_matcher_infos should be _init_req_xgrammar_matcher_infos (grammar instead of grammer).
| def _init_req_xgrammer_matcher_infos(self, run_reqs: List[InferReq]): | |
| def _init_req_xgrammar_matcher_infos(self, run_reqs: List[InferReq]): |
lightllm/server/router/model_infer/mode_backend/chunked_prefill/impl_for_xgrammar_mode.py
Show resolved
Hide resolved
| MAX_ROLLBACK_TOKENS = 200 | ||
|
|
||
|
|
||
| class XGrammarGrammar: |
There was a problem hiding this comment.
Add docstrings to the XGrammarGrammar class and its methods to improve code readability.
…l/impl_for_xgrammar_mode.py Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
1 participant
No description provided.