miniTE aligns native NVFP4 reinforcement learning across engines
Original title: miniTE: Diagnosing and Bridging the Execution Gap in Native NVFP4 Reinforcement Learning
Authors report a Qwen3-30B-A3B learner on one B300 node at 33,540 tokens/s, 1.39× BF16 and 1.47× naive TE NVFP4.
The paper says matching weights and numerical formats is not enough for consistent policy evaluation when RL training and sampling use separate engines. miniTE adds execution diagnostics, cross-engine numerical alignment, and native W4A4 GEMM without changing the RL objective. On Qwen3-30B-A3B, authors report stable long-horizon training with benchmark quality close to BF16 and better than naive Transformer Engine NVFP4. They released the framework, kernels, diagnostics, and training configurations.
Unconfirmed: Throughput and quality claims are the authors' measurements, not independently verified.
- Optimized learner reaches 33,540 tokens/s on one NVIDIA B300 GPU node.
- Throughput is 1.39× BF16 and 1.47× TE NVFP4 for complete learner updates.
- miniTE uses native W4A4 GEMM and does not change the RL objective.
- Authors released kernels, diagnostics, and training configurations.
- Vendors
- NVIDIA
- AI models
- Qwen3-30B-A3B
From the original article
1 min read at huggingface.co
Native NVFP4 promises to accelerate reinforcement learning (RL) for large language models (LLMs), but achieving stable learning across separate training and sampling engines remains challenging. We show that matching model weights and numerical formats is insufficient to ensure consistent policy evaluation, highlighting the importance of cross-engine numerical execution.
By Zhen Li, Mingfa Feng, Yiming Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.