[WIP] MLP weight prefetching unification #15482
Conversation
Summary of ChangesHello @terfendail, 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 refactors the MLP weight prefetching mechanism, consolidating the NPU-specific prefetching logic into the model loader. This change aims to provide a unified and generic approach to weight prefetching across various models, enhancing NPU inference performance and simplifying future model integrations. It also introduces a new configuration option to control this feature. Highlights
🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console. 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request refactors the MLP weight prefetching logic for NPUs by centralizing it in the model loader. This is a good architectural improvement, promoting code reuse and simplifying model-specific implementations. The changes in qwen3.py correctly remove the now-redundant local prefetching logic, and the new server argument is added cleanly.
My review focuses on the new implementation in loader.py. I've identified a critical bug in how attributes are accessed and some opportunities to improve code clarity and maintainability. Specifically, there's a typo using a non-existent method, some commented-out code that should be removed, and a suggestion to make the MLP layer name configuration more robust.
| def mlpwrap(fwd): | ||
| @wraps(fwd) | ||
| def mlpwrapper(*args, **kwds): | ||
| fwdres = fwd(*args, **kwds) | ||
| if _is_npu and get_cmo_stream(): | ||
| wait_cmo_stream() | ||
| return fwdres | ||
| return mlpwrapper | ||
| lmlp.forward = mlpwrap(lmlp.forward) |
There was a problem hiding this comment.
The use of a nested function mlpwrap to create the mlpwrapper decorator is a bit verbose and can be simplified for better readability. You can achieve the same result more directly by capturing the original forward method and then defining and assigning the wrapper.
| def mlpwrap(fwd): | |
| @wraps(fwd) | |
| def mlpwrapper(*args, **kwds): | |
| fwdres = fwd(*args, **kwds) | |
| if _is_npu and get_cmo_stream(): | |
| wait_cmo_stream() | |
| return fwdres | |
| return mlpwrapper | |
| lmlp.forward = mlpwrap(lmlp.forward) | |
| original_forward = lmlp.forward | |
| @wraps(original_forward) | |
| def mlp_wrapper(*args, **kwds): | |
| result = original_forward(*args, **kwds) | |
| if _is_npu and get_cmo_stream(): | |
| wait_cmo_stream() | |
| return result | |
| lmlp.forward = mlp_wrapper |
terfendail
force-pushed
the
prefetch_mlp
branch
from
bdfc766 to
a1fa036
Compare
2 participants
Motivation
Weight prefetching could speedup model inference on NPU
Modifications
Move weight prefetch configuration to model loader for unification between different models to support weight prefetching on any model
Accuracy Tests
Benchmarking and Profiling
Checklist