[MXFP4] Move out of experimental folder

[dsikka]

SUMMARY:

• Move out of experimental as supported in vLLM as of: [MXFP4] Support for linear layers + compressed-tensors integration vllm#41664

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 8382ebde-5f47-4f0a-a096-68b1bb2fc9ee

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

Walkthrough

Documentation links referencing MXFP4 quantization examples are redirected from the experimental/mxfp4/ directory to examples/quantization_w4a4_mxfp4/. The experimental directory's MXFP4 README is removed.

Changes

MXFP4 Example Documentation Reorganization

Layer / File(s)	Summary
Documentation References README.md, examples/autoround/README.md	Links to MXFP4 examples updated from experimental/mxfp4/ to examples/quantization_w4a4_mxfp4/.
Deprecated Documentation experimental/mxfp4/README.md	MXFP4 Quantization README section removed as examples have been consolidated into primary examples directory.

Estimated code review effort

🎯 1 (Trivial) | ⏱️ ~3 minutes

Suggested labels

enhancement, w4a16

Suggested reviewers

• brian-dellabetta

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely describes the main change: moving MXFP4 out of the experimental folder, which aligns with the file modifications and PR objectives.
Description check	✅ Passed	The description directly relates to the changeset by explaining the reason for moving MXFP4 out of experimental (vLLM support), matching the changes to documentation links and folder reorganization.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

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

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch move_mxfp4

---

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.}

[gemini-code-assist]

Code Review

This pull request migrates MXFP4 quantization examples from the experimental directory to the main examples directory. The review feedback focuses on maintaining consistency and documentation quality: it suggests removing an inconsistent trailing slash in the README, aligning relative paths in the AutoRound documentation, and updating the 'What's New!' and 'Known Issues' sections to reflect current support status. Additionally, it is recommended to move and update the existing README rather than deleting it to ensure the new directory remains well-documented.

[dsikka]

@coderabbitai the w4a16 label is incorrect. this is w4a4

[coderabbitai]

Tip

For best results, initiate chat on the files or code changes.

🧩 Analysis chain

🏁 Script executed:

#!/bin/bash
# Search for w4a16 in README and examples related to mxfp4
rg -n "w4a16" README.md examples/ experimental/ --include="*.md" -C2

Repository: vllm-project/llm-compressor
Length of output: 442

---

🏁 Script executed:

#!/bin/bash
rg -n "w4a16" README.md examples/ experimental/ -g "*.md" -C2
# Also check broader context
rg -n "w4a16" -g "*.md" -C2

Repository: vllm-project/llm-compressor
Length of output: 16044

---

🏁 Script executed:

#!/bin/bash
gh pr diff 2685 --repo vllm-project/llm-compressor

Length of output: 76

---

🏁 Script executed:

#!/bin/bash
echo "=== README.md Weight and Activation section ==="
grep -n -A2 -B2 "Weight and Activation\|w4a4_mxfp4\|MXFP4.*w4a\|w4a.*MXFP4\|w4a.*fp4\|fp4.*w4a" README.md

echo ""
echo "=== autoround README MXFP4 section ==="
grep -n -A2 -B2 "MXFP4\|mxfp4\|w4a4\|w4a16" examples/autoround/README.md

Repository: vllm-project/llm-compressor
Length of output: 6444

---

🏁 Script executed:

#!/bin/bash
# Get the diff between main and the PR branch
git diff origin/main...HEAD -- README.md examples/autoround/README.md experimental/mxfp4/README.md

Repository: vllm-project/llm-compressor
Length of output: 41

---

🏁 Script executed:

#!/bin/bash
git log --oneline -5
echo "=== branches ==="
git branch -a | head -20
echo "=== show changed files ==="
git show --stat HEAD
echo "=== show README diff ==="
git show HEAD -- README.md examples/autoround/README.md experimental/mxfp4/README.md

Repository: vllm-project/llm-compressor
Length of output: 61812

---

@dsikka You're absolutely right, my apologies for the error! The examples/quantization_w4a4_mxfp4/ path clearly indicates w4a4 — both weights and activations are quantized to 4-bit FP (MXFP4). That's distinct from the weight-only examples/quantization_w4a16_fp4/mxfp4/ path (w4a16). My review comment was incorrect.

---

✏️ Learnings added

Learnt from: dsikka
URL: https://github.com/vllm-project/llm-compressor/pull/2685

Timestamp: 2026-05-05T19:25:06.162Z
Learning: In the vllm-project/llm-compressor repository, MXFP4 with activation quantization lives at `examples/quantization_w4a4_mxfp4/` and is a **w4a4** scheme (both weights and activations quantized to 4-bit FP). The separate weight-only MXFP4 path is `examples/quantization_w4a16_fp4/mxfp4/` (w4a16). Do not confuse the two when reviewing MXFP4-related changes.

Note: Learnings are effective only in the context of similar code segments. To apply general coding guidelines, please configure review instructions. You can manage existing learnings in the app.

[brian-dellabetta]

👍