Hugging Face (Twitter)

RT @ben_burtenshaw: TIL: you can scale rl environments as hub spaces to batch sizes of 128 on free hardware! This is ideal for most (humble) RL workloads. Plus, openenv scales from there.

this weekend I'm running a benchmarking experiment on openenv RL environments to see what scale they can reach on the main hardware types: (free/ paid) hub spaces, local docker, local python, and slurm+envoy.

the idea is to have a single env definition (openenv) which takes us from hacker scale to hyper scale, without the env getting in the way.

we're sharing this in detail on wednesday in GPU-Mode