miniTE Brings Stable Native NVFP4 RL, 1.39x Faster Than BF16 on B300 GPUs
InfiX-ai · hf · 2026-10-27
InfiX.ai and NVIDIA release miniTE, a lightweight framework that diagnoses and bridges the execution gap in native NVFP4 reinforcement learning for LLMs.
- Key finding: matching model weights and numerical formats across separate training and sampling engines is insufficient — cross-engine numerical alignment is what ensures consistent policy evaluation.
- miniTE combines execution diagnostics, cross-engine numerical alignment, and efficient native W4A4 GEMM computation, improving consistency without modifying the RL objective.
- Long-horizon RL on Qwen3-30B-A3B trains stably with evaluation quality close to BF16, outperforming naive Transformer Engine NVFP4.
- On a single 8×B300 GPU node, the optimized learner reaches 33,540 tokens/s — 1.39× BF16 and 1.47× TE NVFP4 throughput for complete learner updates.
Framework, kernels, diagnostics, and training configs are open-sourced on GitHub.
More from Infra
- YC-backed Isoquant claims 1.2T tokens and 1,000+ customers in its first month — ycombinator · 2026-10-27
- Off-grid data center gas power emits 63-80% more than the grid — 2C_ornot2C · 2026-10-27
- Runware puts FFmpeg in the cloud: common commands via API from $0.02/GB-second — aziz4ai · 2026-10-27
- Google hiring anti-scraper analyst at up to $216K/year to guard search results from LLMs — gaganghotra_ · 2026-10-27
- Kimi K3 on AMD MI355X: +38% agentic throughput, +86% peak vs NVIDIA B300 — YouJiacheng · 2026-10-27
- SemiAnalysis: Anthropic S-1 shows inference is great business, neoclouds still thrive — johncoogan · 2026-10-27