Dev feature caching #68
Conversation
…reshold of taylorseer
![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 @Linboyan-trc, 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 introduces a comprehensive framework for implementing and utilizing different feature caching strategies within the transformer inference process for both Hunyuan and Wan models. The goal is to potentially improve inference speed by selectively skipping full computations and using cached approximations based on various algorithms (Tea, Taylor, Ada, Custom).
Highlights
- Feature Caching Framework: Introduced a base class (
BaseTransformerInfer) and refactored existing transformer inferrers (HunyuanTransformerInfer,WanTransformerInfer) into base classes (BaseHunyuanTransformerInfer,BaseWanTransformerInfer) and specific implementations to support various feature caching strategies. - New Caching Strategies: Added three new feature caching implementations: Taylor (using Taylor series approximation on intermediate block outputs), Ada (employing an adaptive skipping mechanism based on residual changes), and Custom (a different residual caching approach).
- Scheduler Integration: Updated base scheduler (
BaseScheduler) to manage caching state (caching_records,step_index) and added caching-specific scheduler classes (HunyuanSchedulerTeaCaching,HunyuanSchedulerTaylorCaching,HunyuanSchedulerAdaCaching,HunyuanSchedulerCustomCaching,WanSchedulerTeaCaching,WanSchedulerTaylorCaching,WanSchedulerAdaCaching,WanSchedulerCustomCaching) to coordinate with the new transformer inferrers. - Configuration and Scripts: Added new configuration files (
configs/caching/*.json) and inference scripts (scripts/run_*.sh) to enable and demonstrate the new Ada and Custom caching strategies, and updated existing ones for Taylor and Tea caching.
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 pull request significantly refactors the feature caching system by introducing base classes and specialized implementations for different caching strategies (Tea, Taylor, Ada, Custom) for both Hunyuan and WAN models. The changes improve modularity, consistency, and extensibility. Configuration files and run scripts have been updated accordingly. Key areas of improvement include standardized naming, clearer logic for caching decisions, and consistent resource management. Some minor points like a persistent typo and a commented-out import have been noted.
| def __init__(self, config): | ||
| super().__init__(config) | ||
| # 1. fixed args | ||
| self.decisive_double_block_id = 10 |
There was a problem hiding this comment.
Consider adding a comment to explain the origin of decisive_double_block_id = 10.
| self.moreg_steps = [int(0.1 * config.infer_steps), int(0.9 * config.infer_steps)] | ||
| self.moreg_hyp = [0.385, 8, 1, 2] | ||
| self.mograd_mul = 10 | ||
| self.spatial_dim = 3072 |
There was a problem hiding this comment.
Consider adding a comment explaining the source or significance of spatial_dim = 3072.
| super().__init__(config) | ||
|
|
||
| @torch.compile(disable=not CHECK_ENABLE_GRAPH_MODE()) | ||
| def infer(self, weights, img, txt, vec, cu_seqlens_qkv, max_seqlen_qkv, freqs_cis, token_replace_vec=None, frist_frame_token_num=None): |
There was a problem hiding this comment.
The parameter name frist_frame_token_num appears to have a typo and should likely be first_frame_token_num.
| @@ -13,7 +13,8 @@ | |||
| from lightx2v.models.networks.wan.infer.causvid.transformer_infer import ( | |||
| WanTransformerInferCausVid, | |||
| ) | |||
| from lightx2v.models.networks.wan.infer.feature_caching.transformer_infer import WanTransformerInferTeaCaching | |||
|
|
|||
| # from lightx2v.models.networks.wan.infer.feature_caching.transformer_infer import WanTransformerInferTeaCaching | |||
There was a problem hiding this comment.
If TeaCaching is intended to be supported for WanCausVidModel, the import for WanTransformerInferTeaCaching should be reinstated and the necessary integration completed.
希望在 upstream 创建 dev_feature_caching 分支，并将本 PR 合并到该分支。