#!/bin/bash
# Latent + text-encoder caches for the quick run.
# Latents: batch capped at 4 (video VAE VRAM); TE: batch capped at 4 (32B VL encoder + first-frame image).
# Cache TOML has batch_size=8 in [general]; the CLI flag lowers it per stage.
set -Eeuo pipefail
source /venv/main/bin/activate
cd /workspace/projects/musubi-tuner

M=/workspace/models/MiniMax-H3
P=/workspace/projects/h3_loop
CFG="${1:-$P/configs/dataset_quick_cache.toml}"

echo "=== latent cache ($CFG) ==="
python src/musubi_tuner/minimax_h3_cache_latents.py \
  --dataset_config "$CFG" \
  --vae "$M/vae/minimax_h3_video_vae_fp16.safetensors" \
  --device cuda --batch_size 24 --num_workers 16 \
  --skip_existing --keep_cache

echo "=== text encoder cache (task i2va, guidance empty) ==="
python src/musubi_tuner/minimax_h3_cache_text_encoder_outputs.py \
  --dataset_config "$CFG" \
  --text_encoder "$M/text_encoders/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors" \
  --text_encoder_quantization nvfp4_awq \
  --task i2va --cache_guidance_empty \
  --device cuda --batch_size 24 --num_workers 16 \
  --skip_existing --keep_cache

echo "=== cache sizes ==="
du -sh /workspace/cache/quick/* 2>/dev/null