cd /workspace/musubi-tuner || echo "Failed to open directory"

LATENT_BATCH_SIZE=12
LATENT_NUM_WORKERS=6

TEXT_BATCH_SIZE=1
TEXT_NUM_WORKERS=1

mkdir -p logs

latent_pids=()

for i in 0 1 2 3 4 5 6 7; do
  CUDA_VISIBLE_DEVICES=$i python -m musubi_tuner.ltx2_cache_latents \
    --dataset_config "train/ltxxx_gpu${i}.toml" \
    --ltx2_checkpoint "$LTX2_CKPT" \
    --device cuda \
    --vae_dtype bf16 \
    --ltx2_mode av \
    --ltx2_audio_source video \
    --batch_size "$LATENT_BATCH_SIZE" \
    --num_workers "$LATENT_NUM_WORKERS" \
    --skip_existing \
    > "logs/ltx2_cache_latents_gpu${i}.log" 2>&1 &
  latent_pids+=($!)
done

latent_failed=0
for pid in "${latent_pids[@]}"; do
  if ! wait "$pid"; then
    latent_failed=1
  fi
done

if[ "$latent_failed" -ne 0 ]; then
  echo "Latent caching failed on at least one GPU. Check logs/ltx2_cache_latents_gpu*.log"
else
  echo "Latent caching finished successfully on all 8 GPUs. Starting text cache..."

  text_pids=()

  for i in 0 1 2 3 4 5 6 7; do
    CUDA_VISIBLE_DEVICES=$i python -m musubi_tuner.ltx2_cache_text_encoder_outputs \
      --dataset_config "train/ltxxx_gpu${i}.toml" \
      --ltx2_checkpoint "$LTX2_CKPT" \
      --gemma_root "$GEMMA_ROOT" \
      --gemma_load_in_8bit \
      --device cuda \
      --mixed_precision bf16 \
      --ltx2_mode av \
      --batch_size "$TEXT_BATCH_SIZE" \
      --num_workers "$TEXT_NUM_WORKERS" \
      --skip_existing \
      > "logs/ltx2_cache_text_gpu${i}.log" 2>&1 &
    text_pids+=($!)
  done

  text_failed=0
  for pid in "${text_pids[@]}"; do
    if ! wait "$pid"; then
      text_failed=1
    fi
  done

  if[ "$text_failed" -ne 0 ]; then
    echo "Text caching failed on at least one GPU. Check logs/ltx2_cache_text_gpu*.log"
  else
    echo "Latent cache and text cache both completed successfully."
  fi
fi