PufferLib: open-source RL library trains 60M samples/sec on one GPU
jsuarez · x · 2026-11-11
Joseph Suarez (Neural MMO) says his small-model RL stack is pure C/CUDA C, hitting up to 60M training samples per second on a single GPU. PufferLib is a free, open-source RL library that trains tiny super-human models in seconds, with in-house algorithms (PPO variants), Protein/CARBS hyperparameter tuning, and custom simulation methods. RLC 2025 Best Paper; extensive engineering blog series; commercial support available.
More from Infra
- Rumor: DeepSeek to pretrain ~2T-param V5 on Huawei Ascend 950, release may slip to 2027 — teortaxesTex · 2026-11-11
- WSJ: Masayoshi Son Over-Reserved a Big Chunk of the Gas Turbine Market in AI Frenzy — mattwbaker · 2026-11-11
- FT: Commerce Dept stalls Megaspeed export blacklist despite 4 agencies pushing over Nvidia chip smuggling — rohanpaul_ai · 2026-11-11
- Run Pi Coding Agent Locally with TensorSharp + Qwen3.8-27B: Full CLI & SDK Tutorial — fuzhongkai · 2026-11-11
- Spotify's viral Claude Code token trick tested: the real ~90% savings, numbers and prompt inside — aakashgupta · 2026-11-11
- One Opus run cost $1.99 — 94% went to reading files, not writing code — aakashgupta · 2026-11-11