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

PROJECT_DIR=/data/DAIDASEN/ViSA/CAVI-ReID
PYTHON_BIN=/data/conda/cavi_clip/bin/python
OUTPUT_DIR="$PROJECT_DIR/logs/CARGO/CLIP_CAVI_JOINT_60E_B64_V2"
LAUNCH_LOG="$PROJECT_DIR/logs/CARGO/clip_cavi_any_gpu_launcher.log"

mkdir -p "$OUTPUT_DIR" "$(dirname "$LAUNCH_LOG")"
echo "$(date -Is) waiting for physical GPU 1, 2, 3, or 4" >> "$LAUNCH_LOG"

while true; do
  selected_gpu=""
  for gpu in 1 3 4 2; do
    used_mib=$(nvidia-smi --query-gpu=memory.used --format=csv,noheader,nounits -i "$gpu" | tr -d ' ')
    if [[ "$used_mib" =~ ^[0-9]+$ ]] && (( used_mib < 2000 )); then
      selected_gpu="$gpu"
      break
    fi
  done
  [[ -n "$selected_gpu" ]] && break
  sleep 30
done

echo "$(date -Is) GPU $selected_gpu available; starting CLIP+CAVI" >> "$LAUNCH_LOG"
cd "$PROJECT_DIR"
export CUDA_VISIBLE_DEVICES="$selected_gpu"
export FASTREID_DATASETS=/data/DAIDASEN/datasets
exec "$PYTHON_BIN" baselines/hihr/train.py \
  --cavi-clip \
  --cavi-moe \
  --clip-model /data/DAIDASEN/models/clip-vit-base-patch16 \
  --config-file configs/CARGO/clip_cavi.yml \
  --output-dir "$OUTPUT_DIR" \
  --epochs 60 \
  --batch-size 64 \
  --num-workers 0 \
  --device cuda:0 \
  --checkpoint-period 5 \
  --log-period 20 \
  >> "$OUTPUT_DIR/train.log" 2>&1