Paper Details Fix For Execution Gap In Nvidia's NVFP4 Training

Nvidia’s NVFP4 format is pitched as a way to speed up reinforcement learning for large language models by running both weights and activations at 4-bit precision, but the paper argues that matching formats across a training engine and a separate sampling engine is not enough on its own. The authors show that even when weights and numerical formats line up, the two engines can still evaluate the same policy differently, which can quietly destabilize a long reinforcement learning run before anyone notices the drift.

miniTE is the framework the authors built to close that gap. It combines execution diagnostics that surface where the two engines disagree, a cross-engine alignment step that pulls their numerical behavior back into step, and a native W4A4 GEMM kernel so the low-bit computation itself stays efficient. None of this changes the reinforcement learning objective being optimized. The stated goal is to let a team keep the stability of a higher precision run while still getting the speed that native low-bit math is supposed to deliver, and to do it with less engineering overhead than hand-tuning each engine separately.

The team tested the approach on long-horizon reinforcement learning runs using the Qwen3-30B-A3B model, reporting training stability and benchmark scores close to BF16, and better evaluation quality than a naive NVFP4 setup running on Nvidia’s Transformer Engine. The paper states that the framework, its kernels, the diagnostics tooling and the training configurations have been released, which lets other groups reproduce the comparison rather than take the throughput numbers on faith.

The numbers

• The optimized learner reached 33,540 tokens per second on a single Nvidia B300 GPU node for complete learner updates.
• That is 1.39 times the throughput of a BF16 baseline on the same hardware.
• It is 1.47 times the throughput of the earlier, naive NVFP4 setup on Transformer Engine.

Those multipliers describe one model family on one GPU generation, so they are a starting point rather than a guarantee that every reinforcement learning workload sees the same lift. Still, for anyone sizing a training cluster around Nvidia’s newest silicon, a documented route to stable native low-bit reinforcement learning changes the calculation between buying more BF16 capacity and betting on 4-bit throughput instead.