harness/scripts/run_rl.sh

#!/bin/bash
set -uo pipefail
export PRIME_API_KEY="$(cat "$AGENTPTB_PRIME_KEY_FILE")"
export TMPDIR=/var/lib/agentptb-cache/a/tmp
export PYTHONPATH="/mnt/pvc/users/simon/agentptb/runs/a-opus-max/workspace/pkgs${PYTHONPATH:+:$PYTHONPATH}"
export CUDA_VISIBLE_DEVICES=${RL_GPUS:-2,3}
export WANDB_MODE=disabled
# Installs the render-parity patch in EVERY process this launcher spawns (pkgs/sitecustomize.py
# reads it). It has to be process-wide: the renderer is built in the orchestrator, the harness is
# imported in the env server. Evaluation never sets this, so eval is untouched.
export VF_RENDER_PARITY=1
export VLLM_ENGINE_READY_TIMEOUT_S=2400
# /app holds a DIFFERENT prime-rl build: its orchestrator schema disagrees with this one
# (max_inflight_rollouts vs max_inflight_episodes, train.env vs train.source) and its verifiers
# has no broker runtime at all. The default PATH puts /app/.venv first, so the launcher spawns
# its subprocesses from that build and the generated config fails to validate. Put this venv first.
export PATH="/root/work/a/prime-rl/.venv/bin:$PATH"
cd /root/work/a/prime-rl
exec .venv/bin/rl @ "$@"