[Bugfix] Fix the FP8 kv cache accuracy issue in flashinfer TRT-LLM backend

[elvischenv]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

To fix the FP8 kv cache accuracy issue in flashinfer TRT-LLM backend(#19825).

Test Plan

Check the accuracy with lm_eval.

Test Result

Before

vllm (pretrained=nvidia/Llama-4-Scout-17B-16E-Instruct-FP8,quantization=modelopt,tensor_parallel_size=2,max_model_len=1024,kv_cache_dtype=auto,trust_remote_code=True), gen_kwargs: (temperature=0.0), limit: 500.0, num_fewshot: 5, batch_size: 200
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     5|exact_match|↑  |0.684|±  |0.0208|
|     |       |strict-match    |     5|exact_match|↑  |0.652|±  |0.0213|

After

WIP

(Optional) Documentation Update

[github-actions]

👋 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 fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @elvischenv, 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 focuses on resolving an accuracy and functional bug related to FP8 KV cache within the FlashInfer TRT-LLM backend. The changes ensure that FP8 KV cache operates correctly and efficiently, leading to improved model accuracy and enabling previously unsupported configurations.

Highlights

• Bug Fix: FP8 KV Cache Accuracy: I've addressed a critical accuracy issue with FP8 KV cache when using the FlashInfer TRT-LLM backend. This fix significantly improves the gsm8k accuracy for kv_cache_dtype=auto from 0.690 to 0.932 and resolves a functional issue where kv_cache_dtype=fp8_e4m3 was previously unsupported.
• Enhanced FP8 KV Cache Support: The logic for detecting and enabling TRT-LLM decode attention has been refined. It now explicitly checks for torch.float8_e4m3fn and torch.float8_e5m2 data types, ensuring proper activation of the TRT-LLM path for FP8 KV cache.
• Code Clarity and Consistency: I've refactored the kv_cache_dtype parameter in use_trtllm_decode_attention and renamed data_type to kv_data_type within the FlashInferMetadata class and its usages. This change improves type correctness and clarifies that the parameter refers to the KV cache's torch.dtype.
• V1 Backend Compatibility: Similar fixes and enhancements have been applied to the v1 attention backend to ensure consistency and correct FP8 KV cache behavior across different versions.

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

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 fixes an FP8 KV cache accuracy issue in the flashinfer TRT-LLM backend by checking the actual FP8 torch.dtype instead of a string value, improving accuracy. It also enables this feature in the v1 engine and includes variable renames for clarity. I've suggested using torch.is_floating_point for a more robust type check.

[nvpohanh]

@elvischenv Could you explain in the PR description about why we must use self.runner.cache_config.cache_dtype instead of attn_metadata.kv_data_type?