| set -euo pipefail | |
| RL_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)" | |
| cd "${RL_ROOT}" | |
| export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-0,1,2,3,4,5,6}" | |
| export TOKENIZERS_PARALLELISM=false | |
| export PYTHONUNBUFFERED=1 | |
| export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True | |
| RL_PYTHON="${STATEPLAY_WORLD_PYTHON:-${RL_ROOT}/outputs/rl_posttrain/_envs/world/bin/python}" | |
| IFS=',' read -r -a RL_GPUS <<< "${CUDA_VISIBLE_DEVICES}" | |
| exec "${RL_PYTHON}" -m torch.distributed.run --standalone --nproc_per_node "${#RL_GPUS[@]}" -m rl_posttrain.train_api --config rl_posttrain/configs/api_b_only.yaml "$@" | |