LabscoMCP early access ↗
REPO PACKAGECOMMUNITY

huggingface prime-rl

huggingface · publisherApache-2.0 · Listed as freegithub.com/huggingface/prime-rl

LABSCO SUMMARY

Summary

Three skills that only cover getting PRIME-RL installed, running its inference server, and writing its TOML configs — not the reinforcement-learning training the framework itself is built for.

installation walks through installing prime-rl and its optional extras, such as deep-gemm for FP8 models, and troubleshooting dependency issues. inference-server starts and tests the framework's own inference server — in practice, launching and validating vLLM. toml-config explains how to write and use the TOML configuration files the framework's CLI reads, including overriding individual values from the command line. None of the three touches the SFT or RL trainer, the orchestrator, evals, or multi-node deployment — the parts of PRIME-RL its own readme spends the most words on.

PRIME-RL itself needs at least one NVIDIA GPU — the readme lists RTX 3090 through B200, and points toward renting single GPUs or reserving 1000+ GPU clusters through Prime Intellect's own compute platform for anything larger — so these three skills only help someone who has already decided to run this specific framework on hardware they have or are renting. They won't help evaluate whether PRIME-RL is the right RL framework to pick in the first place.

READ MORE

Not a Hugging Face project, by its own README.

The repository sits under the huggingface GitHub org, but the readme's citation block credits "Prime Intellect" as the author, every CI badge links to PrimeIntellect-ai/prime-rl, and the compute-platform links throughout point at app.primeintellect.ai — nothing in the README ties this project to Hugging Face itself.