[Model] Add ModelConfig class for GraniteMoeHybrid to override default max_seq_len_to_capture #20923
Conversation
…q_len_to_capture Signed-off-by: Thomas Parnell <tpa@zurich.ibm.com>
|
👋 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 @tdoublep, 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 enhances the configuration system by adding a dedicated model configuration class for Granite 4.0 models. The primary goal is to automatically optimize performance for long contexts by ensuring the max_seq_len_to_capture parameter aligns with the model's maximum supported sequence length, thereby improving efficiency and user experience.
Highlights
- Model Configuration: A new
GraniteMoeHybridModelConfigclass has been introduced to provide specific configuration overrides for Granite 4.0 models. - Automatic Sequence Length Optimization: The
max_seq_len_to_captureparameter for GraniteMoeHybrid models is now automatically set tomax_model_len, ensuring that CUDA graph capture covers the full supported sequence length. This addresses performance issues observed with smaller batch sizes and long contexts. - Improved User Experience: This change removes the need for users to manually adjust
max_seq_len_to_capturewhen deploying Granite 4.0 models, providing an optimized default configuration out-of-the-box.
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 ModelConfig class for GraniteMoeHybrid to override the default max_seq_len_to_capture and set it to max_model_len. This ensures optimized configuration for Granite 4.0 models, especially when dealing with long contexts. A potential issue is the unconditional override of max_seq_len_to_capture, which might ignore a user's explicit configuration. This has been raised in the review comments.
| logger.info( | ||
| "Setting max_seq_len_to_capture to %d " | ||
| "to ensure that CUDA graph capture " | ||
| "covers sequences of length up to max_model_len.", | ||
| config.max_model_len) |
There was a problem hiding this comment.
Consider adding a warning log message if max_seq_len_to_capture is being overridden, to inform the user that their specified value is not being used. This will help in debugging and understanding the configuration.
| logger.info( | |
| "Setting max_seq_len_to_capture to %d " | |
| "to ensure that CUDA graph capture " | |
| "covers sequences of length up to max_model_len.", | |
| config.max_model_len) | |
| if hasattr(config, 'max_seq_len_to_capture') and config.max_seq_len_to_capture != config.max_model_len: | |
| logger.warning( | |
| "Overriding user-specified max_seq_len_to_capture to %d " | |
| "to ensure that CUDA graph capture " | |
| "covers sequences of length up to max_model_len.", | |
| config.max_model_len) |
| MODELS_CONFIG_MAP: dict[str, type[VerifyAndUpdateConfig]] = { | ||
| "GteModel": SnowflakeGteNewModelConfig, | ||
| "GteNewModel": GteNewModelConfig, | ||
| "NomicBertModel": NomicBertModelConfig, | ||
| "Qwen3ForSequenceClassification": Qwen3ForSequenceClassificationConfig, | ||
| "XLMRobertaModel": JinaRobertaModelConfig, | ||
| "JinaVLForRanking": JinaVLForSequenceClassificationConfig, | ||
| "GraniteMoeHybridForCausalLM": GraniteMoeHybridModelConfig, |
There was a problem hiding this comment.
Consider adding GraniteMoeHybridForCausalLM to __all__ in vllm/model_executor/models/__init__.py to ensure it is accessible for external use and doesn't break any implicit API contracts.
| @staticmethod | ||
| def verify_and_update_config(vllm_config: "VllmConfig") -> None: | ||
| config = vllm_config.model_config | ||
| config.max_seq_len_to_capture = config.max_model_len |
There was a problem hiding this comment.
The implementation unconditionally sets config.max_seq_len_to_capture = config.max_model_len. Consider checking if config.max_seq_len_to_capture already has a user-defined value before overriding it. This would prevent unexpected behavior if a user has explicitly configured this value.
if not hasattr(config, 'max_seq_len_to_capture'):
config.max_seq_len_to_capture = config.max_model_lenSigned-off-by: Thomas Parnell <tpa@zurich.ibm.com>
github-actions
bot
added
the
ready
|
@DarkLight1337 IDK why TPU test keeps failing (I can't figure it out from the logs) but don't see how it can be related to these changes. |
|
I guess TPU CI failure is fixed by #21007 |
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.Purpose
Granite 4.0 models support long context (e.g., 128k) but right now the default value of
max_seq_len_to_captureis set to 8192. This is repeatedly tripping people up, when trying to do benchmarking on long contexts. We find it really affects the performance for this model when using small batches (probably because the mamba implementation uses many Triton kernels which incur launch overhead).This PR introduces a ModelConfig class to override the default value by setting it to
max_model_len. This change would ensure that anyone deploying Granite 4.0 models will get optimized configuration without needing to manually change things.Test Plan
Test Result
Passes locally.
(Optional) Documentation Update