sk Skill · whitecircle
rl-setup
Wire up online GRPO (RLVR) or environmental (multi-turn tool-use) GRPO for Halo: bring up the separate rollout container — vLLM 0.26.0 (cu13), or SGLang 0.5.17 for environmental GRPO (rollout_backend: sglang, gpt-oss + ep1 only) — point the trainer at it via AsyncTrainingConfig rollout URLs, establish NCCL weight sync through the vendored client in src/distributed/nccl/ (parallelism-aware gather — EP/TP/ETP and…
Open on skills.sh ↗read 2026-09-17
- installs 8w
- 0
- 30-day movement
- starts with the next reading
- Related entries
- 1
- Connections
- 4
Python
- Host repository
- whitecircle/halo
- Allowed tools
- Read, Grep, Glob, Bash
- Host stars
- 20
- Host language
- Python