#!/usr/bin/env bash
# 启动 MiniMax-H3 视频 API。用法: ./api-run.sh [GPU_ID] [PORT] [PARTITION]
# PARTITION=fl2va 收首末帧；ref2va 收任意多张参考图。一个进程一个分区，各占一张卡。
set -uo pipefail
B=/data/models/minimax-h3
GPU="${1:-1}"; PORT="${2:-8010}"; export H3_PARTITION="${3:-fl2va}"
export CUDA_VISIBLE_DEVICES="$GPU"
export DIFFSYNTH_ATTENTION_IMPLEMENTATION=torch
export DIFFSYNTH_MODEL_BASE_PATH="$B/models"
export MODELSCOPE_CACHE="$B/.mscache"
export PYTORCH_ALLOC_CONF=expandable_segments:True
export PYTHONUNBUFFERED=1
cd "$B"
echo "[api] GPU=$GPU PORT=$PORT PARTITION=$H3_PARTITION  首次加载约 240 秒（fp32 权重从机械盘读）"
exec "$B/venv/bin/python" -m uvicorn api:app --host 0.0.0.0 --port "$PORT" --workers 1