#!/bin/bash
#PJM -L rscgrp=b-batch
#PJM -L gpu=1
#PJM -L elapse=4:00:00
#PJM -N precompute_sparse
#PJM -j
#PJM -o logs/precompute_sparse_%j.out

module load cuda/12.2.2
module load cudnn/8.9.7
module load gcc-toolset/12

source /home/pj24002027/ku50002536/Takoai/lfj/lfj/stack_env/bin/activate

# 在 grn_ccfm 目录运行（precompute 脚本依赖 grn_ccfm 的 src/）
cd /home/pj24002027/ku50002536/Takoai/lfj/lfj/GRN/grn_ccfm

export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:256

echo "=========================================="
echo "Job ID:    $PJM_JOBID"
echo "Job Name:  $PJM_JOBNAME"
echo "Start:     $(date)"
echo "Node:      $(hostname)"
echo "GPU:       $(nvidia-smi --query-gpu=name,memory.total --format=csv,noheader 2>/dev/null || echo 'N/A')"
echo "=========================================="

# 输出到 grn_ccfm/cache/ (pca_emb 的两个 shell 脚本引用此路径)
mkdir -p cache

python scripts/precompute_sparse_attn.py \
    --data-name norman \
    --n-top-genes 5000 \
    --fold 1 \
    --split-method additive \
    --topk 30 \
    --use-negative-edge \
    --scgpt-model-dir transfer/data/scGPT_pretrained \
    --max-seq-len 5000 \
    --attn-layer 11 \
    --attn-use-rank-norm \
    --batch-size 2 \
    --top-k 300 \
    --n-pca-pairs 1000 \
    --max-pca-components 64 \
    --output cache/norman_attn_L11_sparse.h5 \
    --device cuda

echo "=========================================="
echo "Finished:  $(date)"
echo "Output:    cache/norman_attn_L11_sparse.h5"
echo "=========================================="