Reduce one step decode for draft model.#11561
Conversation
Summary of ChangesHello @hnyls2002, 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 streamlines the speculative decoding process for draft models by adjusting the iteration count in several attention backend modules. The change uniformly reduces the number of steps by one, aiming to optimize performance or correct an indexing issue related to Highlights
Using Gemini Code AssistThe 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
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 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
|
There was a problem hiding this comment.
Code Review
This pull request correctly reduces the number of draft decoding steps by one across various attention backend implementations. The change from range(self.speculative_num_steps) to range(self.speculative_num_steps - 1) aligns with the logic of speculative decoding, where speculative_num_steps - 1 draft tokens are generated. The fix is applied consistently in most files. However, I've pointed out a potential incomplete fix in trtllm_mha_backend.py where other methods might still contain the original buggy loop range, which could lead to an IndexError.
This PR introduces the DraftBackendFactor class to dispatch the mult-step draft decode backend. In any case, the draft backend only initializes attention backends for
speculative_num_steps - 1times.