Revert "[AMD] Fix RotaryEmbedding crash on AMD/ROCm (regression from #17934)"

[HaiShaw]

Reverts #18903

[gemini-code-assist]

Summary of Changes

Hello @HaiShaw, 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 rolls back a prior attempt to resolve a RotaryEmbedding crash on AMD/ROCm. The reversion removes the forward_hip method, indicating that the previous solution was either ineffective or introduced new issues, necessitating a return to the state before that change.

Highlights

• Revert Previous Fix: This pull request reverts the changes introduced in [AMD] Fix RotaryEmbedding crash on AMD/ROCm (regression from #17934) #18903, which aimed to address a RotaryEmbedding crash on AMD/ROCm.
• RotaryEmbedding Changes: The forward_hip method, specifically designed for HIP/ROCm implementation, has been removed from the RotaryEmbedding class.

🧠 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.

Changelog

• python/sglang/srt/layers/rotary_embedding.py
  • Removed the forward_hip method.

Activity

• No specific activity (comments, reviews, progress) has been recorded for this pull request yet.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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. 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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request reverts the fix for the RotaryEmbedding crash on AMD/ROCm platforms. The removal of the forward_hip method forces standard models to use the forward_cuda path, which relies on JIT kernels that have a known dependency on nvidia-smi via tvm_ffi. This will likely re-introduce crashes for ROCm users unless the underlying JIT kernel loading mechanism has been resolved in a separate, non-visible change. Subclasses like MRotaryEmbedding and DeepseekScalingRotaryEmbedding have their own workarounds, but the base class remains vulnerable.

I am having trouble creating individual review comments. Click here to see my feedback.

python/sglang/srt/layers/rotary_embedding.py (392-403)

Removing the forward_hip implementation re-introduces a potential crash on AMD/ROCm GPUs. As noted in the removed docstring, the JIT kernels used in the forward_cuda fallback path depend on tvm_ffi, which invokes nvidia-smi to detect CUDA compute capability. This fails on AMD hardware. By reverting this fix, standard models (like Llama) using the base RotaryEmbedding class will attempt to use these JIT kernels on ROCm and crash. If the JIT kernel loading mechanism hasn't been updated to support ROCm without nvidia-smi, this revert should be reconsidered or replaced with a more robust fix, such as setting self._forward_method = self.forward_native in __init__ for HIP devices (similar to the implementation in DeepseekScalingRotaryEmbedding).