run_ablation.sh

#!/usr/bin/env bash
###############################################################################
# StepProbe — Dataset-size ablation for the restoration step.
#
# Trains a QLoRA restoration adapter at several silver-bullet dataset sizes,
# evaluates each on a primary benchmark, then renders fig_paper_6_ablation.pdf.
#
# Why qwen25-7b + GPTQ w4 and not r1-qwen-7b?
#   The forest plot shows r1-qwen-7b does NOT respond to restoration on math500
#   (ΔAcc ≈ +1 pp, ns) — a flat ablation curve would be a weak paper story.
#   qwen25-7b + GPTQ on math500 is the cell with the strongest significant
#   positive response (+5.4 pp, p<.01), so the ablation curve should show a
#   meaningful upward trend as N grows. Override with env vars if you want
#   different config.
#
# Usage:
#   bash run_ablation.sh                       # defaults
#   MODEL_TAG=r1-qwen-1.5b bash run_ablation.sh  # override model
#   SAMPLE_SIZES='25 50 100 200' bash run_ablation.sh  # override sweep
#
# Expected runtime: ~90-120 min on a single 3090 Ti.
###############################################################################
set -euo pipefail

PROJECT_DIR="$(cd "$(dirname "$0")" && pwd)"
cd "$PROJECT_DIR"

export CUDA_VISIBLE_DEVICES="${CUDA_VISIBLE_DEVICES:-1}"
export HF_HUB_DOWNLOAD_TIMEOUT=300

# ====== Config (override via env) ==========================================
MODEL_HF="${MODEL_HF:-Qwen/Qwen2.5-7B-Instruct}"
MODEL_TAG="${MODEL_TAG:-qwen25-7b}"
QUANT_TAG="${QUANT_TAG:-gptq_w4}"
BENCHMARK="${BENCHMARK:-math500}"
SAMPLE_SIZES_STR="${SAMPLE_SIZES:-50 100 250 500}"
IFS=' ' read -r -a SAMPLE_SIZES <<< "$SAMPLE_SIZES_STR"

PY="${PY:-python}"
MAX_TOKENS="${MAX_TOKENS:-4096}"
# vLLM will OOM if another process (e.g. ollama) is already holding a chunk
# of the GPU. 0.55 gives a 7B NF4 model plenty while leaving ~11 GB free for
# anything else — override with GPU_MEM=... on GPUs where more is free.
GPU_MEM="${GPU_MEM:-0.55}"

# ====== Paths ==============================================================
ABLATION_ROOT="${PROJECT_DIR}/results/ablation/${MODEL_TAG}_${QUANT_TAG}"
DIAG_DIR="${PROJECT_DIR}/results/diagnosis/${QUANT_TAG}/${MODEL_TAG}"
REF_DIR="${PROJECT_DIR}/results/segmented/fp16/${MODEL_TAG}"
LOG_DIR="${PROJECT_DIR}/logs"
TIMESTAMP=$(date +%Y%m%d_%H%M%S)
LOG_FILE="${LOG_DIR}/ablation_${TIMESTAMP}.log"

mkdir -p "$ABLATION_ROOT" "$LOG_DIR"

# ====== Helpers ============================================================
log() {
    local msg="[$(date '+%H:%M:%S')] $1"
    echo "$msg" | tee -a "$LOG_FILE"
}

run_or_skip() {
    local desc=$1
    local skip_check=$2
    local cmd=$3
    if eval "$skip_check"; then
        log "SKIP: $desc"
    else
        log "RUN: $desc"
        eval "$cmd" 2>&1 | tee -a "$LOG_FILE"
    fi
}

# ====== Preflight ==========================================================
log "=============================================="
log "StepProbe restoration ablation"
log "  Model:      $MODEL_HF ($MODEL_TAG)"
log "  Quant:      $QUANT_TAG"
log "  Benchmark:  $BENCHMARK"
log "  Sample Ns:  ${SAMPLE_SIZES[*]}"
log "  Output:     $ABLATION_ROOT"
log "  Log:        $LOG_FILE"
log "=============================================="

# Prerequisites: need diagnosis + FP16 reference segmentation for this config.
if [[ ! -d "$DIAG_DIR" ]]; then
    log "ERROR: missing diagnosis dir: $DIAG_DIR"
    log "       run 'bash run_all.sh --phase 5' to generate it first"
    exit 1
fi
if [[ ! -d "$REF_DIR" ]]; then
    log "ERROR: missing FP16 reference segmentation: $REF_DIR"
    log "       run 'bash run_all.sh --phase 4' to generate it first"
    exit 1
fi

# ====== Main sweep =========================================================
for N in "${SAMPLE_SIZES[@]}"; do
    log ""
    log "=============================================="
    log "Ablation N=$N"
    log "=============================================="

    CFG_DIR="${ABLATION_ROOT}/n${N}"
    ADAPTER="${CFG_DIR}/qlora/adapter"
    MERGED="${CFG_DIR}/qlora/merged_fp16"
    INF_DIR="${CFG_DIR}/inference"
    OUT_FILE="${INF_DIR}/${BENCHMARK}_run0.jsonl"

    # --- 1. Restoration training --------------------------------------------
    run_or_skip "QLoRA restoration (N=$N)" \
        "[[ -f '${ADAPTER}/adapter_model.safetensors' ]]" \
        "$PY -m stepprobe.restore \
            --model '$MODEL_HF' \
            --diagnosis '$DIAG_DIR' \
            --ref '$REF_DIR' \
            --output '$CFG_DIR' \
            --method qlora \
            --max-samples $N \
            --epochs 3 \
            --lr 2e-4 \
            --batch-size 4"

    # --- 2. Merge + inference (skip entirely if inference output already exists) ----
    if [[ -f "$OUT_FILE" ]]; then
        log "SKIP: vLLM inference on $BENCHMARK (N=$N) — $OUT_FILE exists"
    else
        run_or_skip "Merge adapter → FP16 (N=$N)" \
            "[[ -f '${MERGED}/config.json' ]]" \
            "$PY ${PROJECT_DIR}/scripts/merge_adapter.py \
                --model '$MODEL_HF' \
                --adapter '$ADAPTER' \
                --output '$MERGED'"

        $PY -c "import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None" 2>&1 | tee -a "$LOG_FILE"

        run_or_skip "vLLM inference on $BENCHMARK (N=$N)" \
            "[[ -f '$OUT_FILE' ]]" \
            "$PY ${PROJECT_DIR}/scripts/run_inference.py \
                --model '$MERGED' \
                --quant bnb_nf4 \
                --bits 4 \
                --benchmark $BENCHMARK \
                --output '$INF_DIR' \
                --max-tokens $MAX_TOKENS \
                --num-runs 1 \
                --gpu-memory-utilization $GPU_MEM"
    fi

    # --- 4. Segmentation + diagnosis on this ablation config ---------------
    SEG_DIR="${CFG_DIR}/segmented"
    DIAG_OUT="${CFG_DIR}/diagnosis"
    run_or_skip "Segment (N=$N)" \
        "[[ -f '${SEG_DIR}/${BENCHMARK}_run0.jsonl' ]]" \
        "$PY -m stepprobe.segment --input '$INF_DIR' --output '$SEG_DIR' --quant '${QUANT_TAG}_ablation_n${N}'"

    run_or_skip "Diagnose vs FP16 (N=$N)" \
        "[[ -f '${DIAG_OUT}/${BENCHMARK}_run0.jsonl' ]]" \
        "$PY -m stepprobe.diagnose --ref '$REF_DIR' --hyp '$SEG_DIR' --output '$DIAG_OUT' --alignment dtw"

    # --- 5. Free disk (merged FP16 is ~14 GB for a 7B model) ---------------
    if [[ -d "$MERGED" ]]; then
        log "Cleaning up merged FP16 dir: $MERGED"
        rm -rf "$MERGED"
    fi

    $PY -c "import torch; torch.cuda.empty_cache() if torch.cuda.is_available() else None" 2>&1 | tee -a "$LOG_FILE"
done

# ====== Build the ablation figure =========================================
log ""
log "=============================================="
log "Rendering fig_paper_6_ablation.pdf"
log "=============================================="
$PY ${PROJECT_DIR}/scripts/make_ablation_figure.py \
    --ablation-root "$ABLATION_ROOT" \
    --model "$MODEL_TAG" \
    --quant "$QUANT_TAG" \
    --benchmark "$BENCHMARK" \
    --metrics "${PROJECT_DIR}/results/metrics" \
    --output "${PROJECT_DIR}/figures/paper/fig_paper_6_ablation.pdf" 2>&1 | tee -a "$LOG_FILE"

log ""
log "=============================================="
log "DONE"
log "  Figure:   figures/paper/fig_paper_6_ablation.pdf"
log "  Log:      $LOG_FILE"
log "=============================================="