fallback to fa2 (instead of fa3) for unsupported configuration (bf16 Q, Fp8 KV)

[saltyminty]

📌 Description

When the KV cache dtype is FP8, Q dtype is BF16, GPU is hopper, and the backend is set to "auto", we should set the backend as "fa2" instead of "fa3".

🔍 Related Issues

#2348

🚀 Pull Request Checklist

Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.

🧪 Tests

• Tests have been added or updated as needed.
• All tests are passing (unittest, etc.).

Reviewer Notes

Summary by CodeRabbit

• Chores
  • Added configuration validation: when using an FP8 key/value cache, queries must also use FP8 to ensure FA3 backend compatibility. This prevents unsupported dtype combinations, surfaces clear validation errors earlier, and avoids runtime failures when mixing FP8 KV caches with non-FP8 queries.

[gemini-code-assist]

Summary of Changes

Hello @saltyminty, 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 addresses an incompatibility issue where the Flash Attention 3 (FA3) backend was being incorrectly selected for certain configurations involving FP8 KV cache and non-FP8 query data types, specifically on Hopper GPUs. By adding a precise check, the system will now correctly fall back to Flash Attention 2 (FA2) in these unsupported scenarios, preventing potential errors and ensuring optimal performance.

Highlights

• FA3 Backend Compatibility: Introduced a new condition in the is_fa3_backend_supported function to ensure that Flash Attention 3 (FA3) is not used when the KV cache data type is FP8 (e.g., torch.float8_e4m3fn or torch.float8_e5m2) but the query data type is not also FP8. This forces a fallback to FA2 for such configurations.

🧠 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

• flashinfer/utils.py
  • Added a check to is_fa3_backend_supported to prevent FA3 usage with FP8 KV cache and non-FP8 query.

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 correctly adds a check to is_fa3_backend_supported to handle an unsupported configuration where the KV cache is FP8 but the query is not, ensuring a fallback to the fa2 backend. The change is logical and addresses the issue described. I have one minor suggestion to improve the readability and consistency of the new conditional logic.

[coderabbitai]

📝 Walkthrough

Walkthrough

The PR tightens FA3 backend validation in flashinfer/utils.py: if KV cache dtype is FP8 (float8_e4m3fn or float8_e5m2), queries must also be FP8; otherwise FA3 is not supported. Existing checks (custom masks, positional encoding mode, FP16 QK reductions) remain.

Changes

Cohort / File(s)	Summary
FA3 Backend Constraints flashinfer/utils.py	Added a conditional disallowing FA3 when dtype_kv is FP8 but dtype_q is not FP8, tightening backend compatibility logic.
Repository metadata manifest_file, pyproject.toml, requirements.txt	Minor metadata/packaging edits listed in the diff (non-functional).

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Possibly related PRs

• test: Mark test_fp8_prefill.py as xfail on SM90 #2038 — Addresses FA3/FA2 selection for FP8/mixed-precision attention; related to FP8 backend checks.

Suggested reviewers

• nvmbreughe

Poem

🐰 I nudge the bytes and tap the keys,

FP8 dances with matching pleas,
KV and Q now hop in pair,
FA3 hums through cotton air,
A carrot code fix — light as air.

🚥 Pre-merge checks | ✅ 3

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately describes the main change: adding a constraint that falls back to FA2 backend when using BF16 queries with FP8 KV cache, which directly matches the changeset.
Description check	✅ Passed	The description includes the required Description and Related Issues sections with specific context. However, the PR Checklist sections are not completed (checkboxes unchecked), which are part of the template.
Docstring Coverage	✅ Passed	Docstring coverage is 100.00% which is sufficient. The required threshold is 80.00%.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch fix/mingyangw/2348

---

No actionable comments were generated in the recent review. 🎉

Tip

Issue Planner is now in beta. Read the docs and try it out! Share your feedback on Discord.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[yongwww]

@saltyminty please rebase the PR onto the latest main to kick off CI.

[saltyminty]

/bot run

[flashinfer-bot]

GitLab MR !310 has been created, and the CI pipeline #43829833 is currently running. I'll report back once the pipeline job completes.

[flashinfer-bot]

[FAILED] Pipeline #43829833: 15/20 passed