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

# Usage:
#   HF_HOME=/data/adaptai/hf_cache \
#   ./scripts/vllm_launch.sh /data/models/Qwen/Qwen3-4B-Instruct-2507 qwen3_4b_ins_2507 18083 "--gpu-memory-utilization 0.90 --trust-remote-code"

MODEL_PATH=${1:?model path}
ALIAS=${2:?served alias}
PORT=${3:?port}
EXTRA_ARGS=${4:-"--gpu-memory-utilization 0.90 --trust-remote-code"}

HF_HOME=${HF_HOME:-/data/adaptai/hf_cache}
TRANSFORMERS_CACHE=${TRANSFORMERS_CACHE:-$HF_HOME}

CMD=(/usr/bin/python3 -m vllm.entrypoints.openai.api_server \
  --model "$MODEL_PATH" \
  --served-model-name "$ALIAS" \
  --host 127.0.0.1 --port "$PORT" \
)

# shellcheck disable=SC2206
CMD+=( $EXTRA_ARGS )

echo "[vllm] starting $ALIAS on :$PORT using $MODEL_PATH"
env HF_HOME="$HF_HOME" TRANSFORMERS_CACHE="$TRANSFORMERS_CACHE" nohup "${CMD[@]}" >"/tmp/vllm-$ALIAS.log" 2>&1 &
sleep 2
pgrep -fl "openai.api_server.*--served-model-name $ALIAS" || true