[BugFix] Fix: ImportError when building on hopper systems #20513
Conversation
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.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 @LucasWilkinson, 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 primarily addresses a critical ImportError that prevented successful builds on Hopper systems by correcting the conditional registration of a specific CUTLASS kernel. It refactors the PyTorch operator binding mechanism for this kernel to align with its conditional compilation, ensuring symbol availability. Additionally, it includes an administrative update to code ownership for build-related files.
Highlights
- Bug Fix: ImportError on Hopper Systems: This pull request resolves an
ImportError(specifically, anundefined symbolerror forcutlass_blockwise_scaled_grouped_mm) that occurred when building the project from source on Hopper (SM100) systems. The issue was traced to a mismatch where the kernel was conditionally compiled for SM100 but its PyTorch operator registration was not similarly conditional, leading to a missing symbol at runtime. - Refactor: Conditional PyTorch Op Registration: The core fix involves moving the PyTorch operator registration for
cutlass_blockwise_scaled_grouped_mmfrom the centralcsrc/torch_bindings.cppfile to its specific implementation file,csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu. This ensures that the operator is only registered when its corresponding CUDA implementation is compiled, resolving the symbol lookup error. - Administrative: Code Ownership Update: As a secondary change, the author, LucasWilkinson, has been added as a code owner for
CMakeLists.txt. This aims to improve oversight and proactively catch similar build-related issues in the future.
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 correctly fixes an ImportError by moving the registration of a conditionally compiled CUDA kernel into its source file, ensuring it's only registered when compiled. The changes in ops.h, torch_bindings.cpp, and the .cu file are logical and address the issue. I've added a comment regarding code formatting in the new registration block to improve consistency and maintainability.
csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu
Outdated
Show resolved
Hide resolved
LucasWilkinson
changed the title
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
LucasWilkinson
added
the
ready
…ct#20513) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
…ct#20513) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
…ct#20513) Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Signed-off-by: Jinzhen Lin <linjinzhen@hotmail.com>
2 participants
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.Purpose
Fix:
when building from source on hopper systems.
Caused by #19757 conditionally compiling
csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cubut did not make the linkage conditional.Also adding my self as a CMakeList code owner to get more eyes and hopefully catch these things early (please let me know if you'd prefer the to be a separate PR .
Test Plan
python -m pytest tests/kernels/moe/test_cutlass_grouped_gemm.pystill passesTest Result
python -m pytest tests/kernels/moe/test_cutlass_grouped_gemm.pypasses(Optional) Documentation Update