The interesting part is the diagnosis of quantization-induced training–sampling mismatch, a failure mode most RLHF and post-training pipelines still handle by trial and error. Engineers running low-precision fine-tuning on Blackwell-class GPUs can adopt TRIAGE and the fused grouped GEMM setup to reclaim update throughput. It benchmarks on 8×B300, placing it alongside NVFP4 tooling rather than general-purpose trainer frameworks.
做 RL 训练的朋友看看这篇：不用重写训练框架，只在 RoPE 那条路保留 FP32 中间量就能防崩溃，四个 Qwen 模型都验证过了。
Anyone chasing reproducible LLM outputs — evaluation harness authors, serving teams and quantisation engineers — should treat BF16 vs FP16 as a real source of run-to-run variance, not an implementation detail. The paper offers a low-overhead fix (threshold-triggered FP32 LM head recomputation) and maps when it fails, including FP8 end-to-end serving. The margin-based analysis also gives practitioners a quick diagnostic for whether their own deployments are at risk.
想让你自己训练的Transformer跑得更快？这篇手把手教你用NVIDIA Transformer Engine和FP8，还有代码和GPU跑分对比。
