scripts/run_all_evals.sh

#!/bin/bash
# =============================================================
# Run all 4 methods in parallel on 6 GPUs
# Assumes: training is done, checkpoint at checkpoints/optical/best.pt
# =============================================================

set -e
export HF_HOME=${HF_HOME:-/root/rivermind-data/huggingface}
export HF_HUB_OFFLINE=1
export TRANSFORMERS_OFFLINE=1

N=${1:-50}  # number of samples, pass as argument: ./run_all_evals.sh 485

echo "=== UIPress Evaluation: $N samples ==="

# GPU 0: Baseline (full resolution)
CUDA_VISIBLE_DEVICES=0 python scripts/eval_all.py \
    --method baseline --max_samples $N \
    > logs/eval_baseline.log 2>&1 &

# GPU 1: VisionZip 256 tokens
CUDA_VISIBLE_DEVICES=1 python scripts/eval_all.py \
    --method visionzip --keep_tokens 256 --max_samples $N \
    > logs/eval_visionzip_256.log 2>&1 &

# GPU 2: VisionZip 128 tokens
CUDA_VISIBLE_DEVICES=2 python scripts/eval_all.py \
    --method visionzip --keep_tokens 128 --max_samples $N \
    > logs/eval_visionzip_128.log 2>&1 &

# GPU 3: EfficientUI 60% prune
CUDA_VISIBLE_DEVICES=3 python scripts/eval_all.py \
    --method efficientui --prune_ratio 0.6 --max_samples $N \
    > logs/eval_efficientui_60.log 2>&1 &

# GPU 4: UIPress 256 tokens
CUDA_VISIBLE_DEVICES=4 python scripts/eval_all.py \
    --method uipress --checkpoint checkpoints/optical/best.pt \
    --target_tokens 256 --max_samples $N \
    > logs/eval_uipress_256.log 2>&1 &

# GPU 5: Resolution scaling 256px
CUDA_VISIBLE_DEVICES=5 python scripts/eval_all.py \
    --method resolution --max_pixels 230400 --max_samples $N \
    > logs/eval_resolution_256.log 2>&1 &

echo "All evaluations launched. Monitor with: tail -f logs/eval_*.log"
wait
echo "=== All evaluations complete ==="