Miles v0.1.3 cuts first RL weight update from 99.5s to 3.9s, shares GPUs across train/rollout

hsu_byron · x · 2026-10-24

RadixArk released Miles v0.1.3, an open-source RL training framework focused on lighter weight and more resilient runs:

• Rebuilt P2P transfer cuts the first weight update from 99.5s to 3.9s and pinned memory per sender from 28 GiB to 2 GiB in benchmarks
• Fully async RL can share GPUs between training and rollout, even on a single GPU; the new fused log-prob backend cuts peak GPU memory by 1/3 in long-response tests
• Fault tolerance: a failed cell no longer blocks weight updates to healthy cells; fully async runs can resume
• LoRA now runs in Megatron raw mode with a unified implementation across model families

Repo: github.com/radixark/miles

Original post →

More from Infra

• NVIDIA makes 8-node DGX Spark clusters official: 97.4% scaling efficiency on 900M training — TheZachMueller · 2026-10-24
• One NVIDIA Vera Rubin rack would set you back at least ~$7M — LinusEkenstam · 2026-10-24
• DSF4.1 hits 280 tok/s on a single RTX 6000 — community chases 500 tok/s setups — capetorch · 2026-10-24
• PS5 Pro sells out in Toronto as memory shortage hits consumer hardware — Rasmic · 2026-10-24
• Open-source kapitsa engine boosts Mac local agent throughput 23% with tree speculative decoding — DotaMate · 2026-10-24
• Nvidia compute unlock boosts RTX 3080 20GB prefill by up to 29%, decode unchanged — No-Refrigerator-1672 · 2026-10-24

Infra channel →