#!/usr/bin/env bash
#
# Llama 4 统一容器入口：
#   1. 幂等应用 CoreX vLLM 的 Llama 4 补丁
#   2. 读取 /models/llama4/config.json，自动判断 Scout / Maverick
#   3. 选择对应的 TP / PP / 端口 / served name，再启动 vLLM
#
# AUTO_DETECT=1（默认）时根据模型 config 自动选择；显式设置的
# TP / PP / PORT / SERVED_NAME 会覆盖检测结果。AUTO_DETECT=0 可完全手动。
#
set -euo pipefail

PATCH_SCRIPT="${LLAMA4_PATCH_SCRIPT:-/opt/llama4/apply_patch.py}"
DETECT_SCRIPT="${LLAMA4_DETECT_SCRIPT:-/opt/llama4/detect_model.py}"
VLLM_BIN="${VLLM_BIN:-vllm}"

python3 "${PATCH_SCRIPT}"

MODEL_PATH="${MODEL_PATH:-/models/llama4}"
AUTO_DETECT="${AUTO_DETECT:-1}"
MODEL_DESCRIPTION=""

if [[ ! -f "${MODEL_PATH}/config.json" ]]; then
  echo "ERROR: 模型目录不完整：${MODEL_PATH}/config.json 不存在" >&2
  echo "请确认已挂载模型，例如：" >&2
  echo "  -v /mnt/disk0/models/LLM-Research/Llama-4-Scout-17B-16E-Instruct:/models/llama4:ro" >&2
  echo "  -v /mnt/disk0/models/LLM-Research/Llama-4-Maverick-17B-128E-Instruct-FP8:/models/llama4:ro" >&2
  exit 1
fi

if [[ "${AUTO_DETECT}" == "1" ]]; then
  DETECT_OUTPUT="$(python3 "${DETECT_SCRIPT}" "${MODEL_PATH}")" || exit 1
  eval "${DETECT_OUTPUT}"
  MODEL_KIND="${MODEL_KIND:-${DETECTED_MODEL_KIND}}"
  TP="${TP:-${DETECTED_TP}}"
  PP="${PP:-${DETECTED_PP}}"
  PORT="${PORT:-${DETECTED_PORT}}"
  SERVED_NAME="${SERVED_NAME:-${DETECTED_SERVED_NAME}}"
  MODEL_DESCRIPTION="${DETECTED_DESCRIPTION:-}"
else
  MODEL_KIND="${MODEL_KIND:-unknown}"
fi

HOST="${HOST:-0.0.0.0}"
TP="${TP:-8}"
PP="${PP:-1}"
PORT="${PORT:-8091}"
SERVED_NAME="${SERVED_NAME:-llama4}"
MAX_MODEL_LEN="${MAX_MODEL_LEN:-4096}"
MAX_NUM_SEQS="${MAX_NUM_SEQS:-1}"
GPU_MEM_UTIL="${GPU_MEM_UTIL:-0.90}"
ENFORCE_EAGER="${ENFORCE_EAGER:-1}"
EXTRA_ARGS="${EXTRA_ARGS:-}"

PP_ARGS=()
if (( PP > 1 )); then
  PP_ARGS=(--pipeline-parallel-size "${PP}")
fi

EAGER_ARGS=()
if [[ "${ENFORCE_EAGER}" == "1" ]]; then
  EAGER_ARGS=(--enforce-eager)
fi

echo "============================================================"
echo "Llama 4 (BI-V150 unified image)"
echo "  model kind : ${MODEL_KIND}"
echo "  model      : ${MODEL_PATH}"
echo "  served name: ${SERVED_NAME}"
echo "  listen     : ${HOST}:${PORT}"
echo "  TP / PP    : ${TP} / ${PP}  (total $((TP * PP)) GPUs)"
echo "  max len    : ${MAX_MODEL_LEN}"
echo "  max seqs   : ${MAX_NUM_SEQS}"
echo "  gpu util   : ${GPU_MEM_UTIL}"
echo "  eager      : ${ENFORCE_EAGER}"
echo "  visible    : ${CUDA_VISIBLE_DEVICES:-all}"
echo "  extra      : ${EXTRA_ARGS}"
if [[ -n "${MODEL_DESCRIPTION}" ]]; then
  echo "  detected   : ${MODEL_DESCRIPTION}"
fi
echo "============================================================"

# shellcheck disable=SC2086
exec "${VLLM_BIN}" serve "${MODEL_PATH}" \
  --host "${HOST}" \
  --port "${PORT}" \
  --served-model-name "${SERVED_NAME}" \
  --tensor-parallel-size "${TP}" \
  "${PP_ARGS[@]}" \
  --max-model-len "${MAX_MODEL_LEN}" \
  --max-num-seqs "${MAX_NUM_SEQS}" \
  --gpu-memory-utilization "${GPU_MEM_UTIL}" \
  "${EAGER_ARGS[@]}" \
  ${EXTRA_ARGS}