#!/usr/bin/env bash
# Serve this complete checkpoint with the vLLM layout patch shipped alongside it.
set -euo pipefail
ROOT=$(cd "$(dirname "$0")" && pwd)
export PYTHONPATH="$ROOT/vllm_patch${PYTHONPATH:+:$PYTHONPATH}"
export VLLM_WORKER_MULTIPROC_METHOD=spawn
export VLLM_SSM_CONV_STATE_LAYOUT=DS
if [ -d /usr/lib/x86_64-linux-gnu ]; then
  export LD_LIBRARY_PATH="/usr/lib/x86_64-linux-gnu${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
fi
exec "${VLLM:-vllm}" serve "$ROOT" --served-model-name "${SERVED_NAME:-e384-rl}" \
  --port "${PORT:-8000}" --tensor-parallel-size 1 --dtype bfloat16 \
  --kv-cache-dtype fp8 --quantization modelopt_mixed --mamba-cache-mode align \
  --enable-prefix-caching --enable-chunked-prefill --max-model-len 72000 \
  --max-num-seqs 128 --gpu-memory-utilization 0.90 --reasoning-parser qwen3 \
  --tool-call-parser qwen3_xml --enable-auto-tool-choice "$@"