#!/usr/bin/env bash
set -euo pipefail

MODEL_NAME="${MODEL_NAME:-Qwen/Qwen3.6-27B}"
SERVED_MODEL_NAME="${SERVED_MODEL_NAME:-qwen36-base}"
HOST="${HOST:-0.0.0.0}"
PORT="${PORT:-8000}"
DTYPE="${DTYPE:-bfloat16}"
TP_SIZE="${TP_SIZE:-1}"
MAX_MODEL_LEN="${MAX_MODEL_LEN:-32768}"
VLLM_EXTRA_ARGS="${VLLM_EXTRA_ARGS:---trust-remote-code}"

echo "Serving base model with vLLM:"
echo "  model:        $MODEL_NAME"
echo "  served name:  $SERVED_MODEL_NAME"
echo "  endpoint:     http://$HOST:$PORT/v1"
echo "  tensor para:  $TP_SIZE"
echo

python -m vllm.entrypoints.openai.api_server \
  --model "$MODEL_NAME" \
  --served-model-name "$SERVED_MODEL_NAME" \
  --host "$HOST" \
  --port "$PORT" \
  --dtype "$DTYPE" \
  --tensor-parallel-size "$TP_SIZE" \
  --max-model-len "$MAX_MODEL_LEN" \
  $VLLM_EXTRA_ARGS