#!/bin/bash
# Usage: serve.sh <model_path> [port] [dp] [gpus]
# Serves a checkpoint for evaluation. The two tool-call flags are mandatory (see eval-kit):
# without them every structured-tool-call episode 400s.
set -euo pipefail
MODEL="${1:?model path}"
PORT="${2:-8000}"
DP="${3:-3}"
export CUDA_VISIBLE_DEVICES="${4:-${SERVE_GPUS:-0,2,3}}"
export VLLM_CACHE_ROOT=${VLLM_CACHE_ROOT:-/var/lib/agentptb-cache/a/tmp/vllm}
export TMPDIR=${TMPDIR:-/var/lib/agentptb-cache/a/tmp}
export VLLM_ENGINE_READY_TIMEOUT_S=2400
exec /root/work/a/prime-rl/.venv/bin/vllm serve "$MODEL" \
  --served-model-name policy \
  --host 0.0.0.0 --port "$PORT" \
  --data-parallel-size "$DP" \
  --tensor-parallel-size 1 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder \
  --limit-mm-per-prompt '{"image":0,"video":0}' \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.85 \
  --trust-remote-code