[Quantization] Add compressed-tensors emulations support for NVFP4

[dsikka]

Purpose

• Enable an option to run emulations when running NVFP4 on a non-b200 machine instead of falling back to always use marlin
• Now if the user sets VLLM_USE_NVFP4_CT_EMULATIONS to 1, emulations will be used for nvfp4 as opposed to nvfp4a16

Test Plan

• All exisiting tests pass; setting VLLM_USE_NVFP4_CT_EMULATIONS uses emulations

[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 @dsikka, 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 enhances the flexibility of NVFP4 quantization by allowing users to explicitly enable compressed-tensors emulations via an environment variable. This change is particularly useful for running NVFP4 models on hardware that lacks native Cutlass FP4 support, providing a fallback mechanism to ensure compatibility and continued functionality.

Highlights

• NVFP4 Emulation Control: Introduced an environment variable, USE_NVFP4_CT_EMULATIONS, to explicitly enable compressed-tensors emulations for NVFP4 quantization. This allows users to force emulation even on machines that do not natively support Cutlass FP4.
• Conditional Emulation Execution: Modified the CompressedTensorsW4A4Fp4 scheme's apply_weights method to conditionally invoke run_nvfp4_emulations when the USE_NVFP4_CT_EMULATIONS environment variable is set, bypassing the default weight application logic.
• Scheme Selection Logic: Updated the _get_scheme_from_parts function to consider the USE_NVFP4_CT_EMULATIONS environment variable as an alternative condition for selecting the CompressedTensorsW4A4Fp4 scheme, alongside native Cutlass FP4 support.

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 introduces an emulation mode for NVFP4 compressed tensors, which can be activated via the USE_NVFP4_CT_EMULATIONS environment variable. This is a useful feature for development and testing on non-b200 hardware.

The implementation is mostly straightforward. The key points are:

1. Environment Variable Control: The feature is controlled by os.environ.get("USE_NVFP4_CT_EMULATIONS") == "1".
2. Emulation Path: When enabled, the code calls run_nvfp4_emulations instead of ops.fp4_gemm.

Critical Concern:

• The most significant issue is that the bias term, which is handled in the original ops.fp4_gemm path, appears to be ignored in the new emulation path. This will lead to incorrect results when a bias is present. The emulation should be updated to correctly apply the bias.

Suggestions:

• Please address the bias handling in the emulation path.
• Consider adding more details to the PR description, including:
  • The name of the environment variable (USE_NVFP4_CT_EMULATIONS).
  • A brief explanation of what the emulation does and any known limitations (e.g., performance, precision differences, or if bias handling is intentionally omitted for some reason).
• The PR description mentions a "Test Plan" and "Test Result" section, but they are empty. Please include information about how this change was tested to ensure the emulation's correctness and that it behaves as expected.

Overall, a good addition, but the bias handling needs to be rectified for the emulation to be accurate.