run_evalcmp.sh

#!/bin/bash
# Same-settings A/B/C KL-by-depth eval: baseline vs approach-(A) regressed ckpt.
cd /workspace
export HF_HOME=/workspace/.hf
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
LOG=/workspace/evalCMP.log
echo "===== BASELINE hn_step7750 =====" > $LOG
python3 diag_qwen_ablation.py --ckpt /workspace/hypernet_qwen/hn_step7750.pt \
  --n_samples 24 --chunk_size 64 --raw_window 32 >> $LOG 2>&1
echo "" >> $LOG
echo "===== REGRESS-A hn_regressA_ep40 =====" >> $LOG
python3 diag_qwen_ablation.py --ckpt /workspace/hypernet_qwen_distill/hn_regressA_ep40.pt \
  --n_samples 24 --chunk_size 64 --raw_window 32 >> $LOG 2>&1
echo "ALLDONE" >> $LOG