code/scripts/run_mi_cpu_pipeline.sh

#!/usr/bin/env bash
# MI workshop pipeline on CPU — 3 grokking seeds + 1 standard control.
# Runs sequentially to avoid OpenMP thrash. Survives SSH disconnect via nohup.
set -uo pipefail
cd /home/garima/CausalGrok

LOG=logs/mi_cpu_pipeline.log
echo "[$(date -Iseconds)] MI CPU pipeline starting" >  "$LOG"

run_m1() {
    local run_dir="$1"
    local tag="$2"
    local extra="${3:-}"
    echo "" >> "$LOG"
    echo "[$(date -Iseconds)] === M1 on $tag : $run_dir ===" >> "$LOG"
    python3 -m experiments.mechinterp_m1 \
        --run_dir "$run_dir" \
        --data_root data/wilds \
        --device cpu \
        --max_samples 600 \
        $extra >> "$LOG" 2>&1
    rc=$?
    echo "[$(date -Iseconds)] $tag exit=$rc" >> "$LOG"
}

# Priority 1: grokking trajectories (15 ckpts each)
run_m1 experiments/runs/20260503-135145_camelyon_v2_grokking_n300_s123 grok_s123
run_m1 experiments/runs/20260503-021649_grokking_n300_s42              grok_s42
run_m1 experiments/runs/20260503-135149_camelyon_v2_grokking_n300_s456 grok_s456

# Priority 2: standard control (only final.pt)
run_m1 experiments/runs/20260502-214859_standard_n300_s42  std_s42  --latest_only

echo "" >> "$LOG"
echo "[$(date -Iseconds)] MI CPU pipeline DONE" >> "$LOG"