#!/usr/bin/env bash
# MiniMax-H3 FL2VA NF4 launcher.  usage:  ./run.sh <GPU_ID> [args for t2va.py ...]
#   ./run.sh 1                                  # turbo 4-step, 480x832x124f
#   ./run.sh 1 --no-turbo --steps 50            # full quality
#   ./run.sh 2 --prompt "..." --seed 7
set -euo pipefail
B=/data/models/minimax-h3
GPU="${1:-1}"; shift || true

export CUDA_VISIBLE_DEVICES="$GPU"
export DIFFSYNTH_ATTENTION_IMPLEMENTATION=torch     # sm_75: no flash/sage; pin instead of auto-picking xformers
export DIFFSYNTH_MODEL_BASE_PATH="$B/models"
export MODELSCOPE_CACHE="$B/.mscache"
export PYTHONUNBUFFERED=1
# 22GB 卡上激活与权重贴着上限，碎片会直接压垮它（实测 reserved-but-unallocated 达 2.49GB）
export PYTORCH_ALLOC_CONF=expandable_segments:True
# NOTE: deliberately NOT setting PYTHONNOUSERSITE -- the venv's site-packages already
# precedes ~/.local in sys.path, so our bitsandbytes 0.50.1 shadows the broken 0.45.1 there.

# GPU0/1 sit on NUMA0, GPU2/3 on NUMA1 -- keep the CPU-offloaded weights on the near node.
case "$GPU" in 0|1) NODE=0 ;; 2|3) NODE=1 ;; *) NODE="" ;; esac
BIND=(); if [ -n "$NODE" ] && command -v numactl >/dev/null 2>&1; then BIND=(numactl --cpunodebind="$NODE" --membind="$NODE"); fi

LOG="$B/logs/gpu${GPU}_$(date +%Y%m%d-%H%M%S).log"
echo "[run] GPU=$GPU numa=${NODE:-none} log=$LOG"
"${BIND[@]}" "$B/venv/bin/python" "$B/scripts/t2va.py" "$@" 2>&1 | tee "$LOG"