#!/bin/bash
set -euo pipefail
MODEL="${1:?}"; PORT="${2:-8010}"; GPU="${3:-3}"; UTIL="${4:-0.30}"
export CUDA_VISIBLE_DEVICES=$GPU
export VLLM_CACHE_ROOT=/var/lib/agentptb-cache/a/tmp/vllm
export TMPDIR=/var/lib/agentptb-cache/a/tmp
export VLLM_ENGINE_READY_TIMEOUT_S=2400
exec /root/work/a/prime-rl/.venv/bin/vllm serve "$MODEL" \
  --served-model-name policy --host 0.0.0.0 --port "$PORT" \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --limit-mm-per-prompt '{"image":0,"video":0}' \
  --max-model-len 65536 --gpu-memory-utilization "$UTIL" --trust-remote-code