transfer/code/CCFM/run_cached.sh

#!/bin/bash
#PJM -L rscgrp=b-batch
#PJM -L gpu=4
#PJM -L elapse=24:00:00
#PJM -j
#PJM -S
#PJM -o /home/hp250092/ku50001222/qian/aivc/lfj/transfer/logs/ccfm_cached_%j.out

# Activate environment
source /home/hp250092/ku50001222/qian/aivc/lfj/stack_env/bin/activate

# Navigate to CCFM
cd /home/hp250092/ku50001222/qian/aivc/lfj/transfer/code/CCFM

export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:256

echo "=========================================="
echo "Job ID:    $PJM_JOBID"
echo "Start:     $(date)"
echo "Node:      $(hostname)"
echo "=========================================="

# Run training with accelerate (multi-GPU) + scGPT cache
accelerate launch --num_processes=4 scripts/run_cascaded.py \
    --data-name norman \
    --d-model 128 \
    --nhead 8 \
    --nlayers 4 \
    --batch-size 48 \
    --lr 5e-5 \
    --steps 200000 \
    --fusion-method differential_perceiver \
    --perturbation-function crisper \
    --noise-type Gaussian \
    --infer-top-gene 1000 \
    --n-top-genes 5000 \
    --use-mmd-loss \
    --gamma 0.5 \
    --split-method additive \
    --fold 1 \
    --scgpt-dim 512 \
    --bottleneck-dim 128 \
    --latent-weight 1.0 \
    --choose-latent-p 0.4 \
    --dh-depth 2 \
    --print-every 10000 \
    --result-path ./result \
    --scgpt-cache-path /home/hp250092/ku50001222/qian/aivc/lfj/transfer/code/CCFM/scgpt_cache_norman.h5

echo "=========================================="
echo "Finished:  $(date)"
echo "=========================================="