Quantization: support FP4 quantized models on AMD CDNA2/CDNA3 GPUs

[fengli1702]

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

Add support for running FP4 quantized models on AMD CDNA2 / CDNA3 GPUs via a new quantization backend petit_nvfp4. This enables vLLM to serve models such as Llama-3.1 / Llama-3.3

Test Plan

Verified that ROCm backend without --quantization still runs normally, and automatically selects the petit-kernel FP4 path when loading FP4 quantized models.

Installed optional dependencies via:

pip install -e .[fp4]

Ran test with:

from vllm import LLM, SamplingParams
llm = LLM(model="/path/to/llama-fp4", enforce_eager=True)  # no quantization arg
prompt = "San Francisco is a"
output = llm.generate(prompt, SamplingParams(temperature=0, max_tokens=15))
print(output[0].outputs[0].text)

Confirmed that the FP4 path (petit-kernel) is selected automatically and outputs are generated successfully.

Test Result

Functional test passed: FP4 quantized models load and run inference successfully on ROCm (MI250/MI300 test nodes) without specifying --quantization.

Non-quantized ROCm inference path unaffected.

Verified via log output that petit-kernel module is loaded when using FP4 models.

(Optional) Documentation Update

Added fp4 extras in setup.py for optional dependency installation.

[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]

Code Review

This pull request adds support for FP4 quantized models on AMD GPUs using the petit-kernel. The changes look mostly good, with new quantization configurations and utility functions. However, I found a critical issue in the PetitNvFp4Config.from_config method where it uses direct dictionary access for optional keys, which can lead to a KeyError and crash the application if the keys are not present in the configuration file. I've suggested a more robust implementation for this method to handle missing keys gracefully.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @fengli1702.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @fengli1702.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork