Fix nan value generated after custom all reduce #8532
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @kkHuang-amd, 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!
I've implemented a fix to prevent NaN values from being generated after our custom all-reduce function, which was previously causing failures in the gsm8k accuracy test. The solution involves changing how certain tensors are initialized within the communication process. Instead of allocating uninitialized memory, these tensors are now explicitly set to zero, ensuring a clean state and preventing the propagation of garbage values that could lead to NaNs.
Highlights
- NaN Value Fix: The core issue of NaN values appearing after custom all-reduce operations, which was causing accuracy test failures, has been addressed.
- Tensor Initialization Change: All instances of
torch.emptyandtorch.empty_likeused for tensor allocation within the custom all-reduce communication logic have been replaced withtorch.zerosandtorch.zeros_like. This ensures that all newly created tensors are explicitly initialized to zero, preventing the propagation of uninitialized memory contents that could lead to NaNs.
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 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 or fill out our survey 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. ↩
|
Hi @HaiShaw : Please check this PR for potential issue fixed |
There was a problem hiding this comment.
Code Review
This pull request addresses a critical bug where NaN values were generated by the custom all-reduce function. The root cause was the use of torch.empty and torch.empty_like for creating buffers and output tensors, which does not guarantee initialization and can lead to using uninitialized memory. The fix correctly replaces these calls with torch.zeros and torch.zeros_like, ensuring all tensors are zero-initialized. This is a robust solution that prevents non-deterministic behavior and correctness issues. The changes are consistent, well-targeted, and look good to me.
|
@yizhang2077 pls have a review |
Motivation
Found NaN value generated after custom all reduce function.
It will cause gsm8k accuracy test failed.
Modifications
Change the empty_like and empty to zeros_like and zeros to ensure the tensor value is initialized to zero
Accuracy Test
python3 benchmark/gsm8k/bench_sglang.py --num-questions 2000 --parallel 2000 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [00:51<00:00, 25.58it/s] Accuracy: 0.948 Invalid: 0.000 Latency: 52.566 s Output throughput: 2409.507 token/sChecklist