[Draft]feat(recipe): add NVFP4 QAT recipe for Qwen3-30B W4A16#36
Draft
zhangyimi wants to merge 6 commits intoverl-project:mainfrom
Draft
[Draft]feat(recipe): add NVFP4 QAT recipe for Qwen3-30B W4A16#36zhangyimi wants to merge 6 commits intoverl-project:mainfrom
zhangyimi wants to merge 6 commits intoverl-project:mainfrom
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
This PR adds support for NVFP4 Quantization-Aware Training (QAT) with FSDP, enabling W4A16 (weight-only) quantization during RL training.
What's included
verl/utils/qat/ module: QATLinear (Triton FP4 fake quantization), scale fusion, NVFP4 quantizer, and vLLM dynamic weight loading patches
Recipe scripts and configs for Qwen3-30B-A3B W4A16 (full quantization & FFN-only quantization)
Detailed README with implementation overview and experimental results
Key Results
Validated on Qwen3-8B-Base (Dense) and Qwen3-30B-A3B-Base (MoE): W4A16 QAT achieves training accuracy on par with BF16 baseline, while without QAT the KL divergence explodes and training crashes.
70.3% weight memory reduction on Qwen3-30B-A3B during rollout (56.88 GiB → 16.89 GiB), freeing ~40 GiB for additional KV Cache capacity.
VeRL PR: verl-project/verl#5190
README: https://github.com/zhangyimi/verl-recipe/blob/006aa5dabb8dac1f2369e52c3ad27455b84e7799/qat/README.md