test.sh

#!/usr/bin/env bash
# Run Gumbel experiments for multiple models

###############################################################################
# Config
###############################################################################
N_PROMPTS=5000
MAX_TOKENS=500
GPU_MEM=0.85
SIGMAS="0.001,0.01,0.1,1.0"
SUPPORT_SIZE=1000
MAX_THRESHOLDS=500
MAX_MODEL_LEN=8192

# Robust HF downloads (avoid git-xet 500s)
export HF_HUB_ENABLE_GIT_XET=0
export GIT_XET=0
export HF_HUB_ENABLE_HF_TRANSFER=1
export HF_HUB_DISABLE_TELEMETRY=1

# Nice-to-have: make Python fail fast on OOMs
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512

# Timestamped sweep dir
TIMESTAMP=$(date +"%Y%m%d_%H%M%S")
SWEEP_DIR="multi_model_sweep_${TIMESTAMP}"
mkdir -p "$SWEEP_DIR"

echo "========================================================================"
echo "Running multi-model Gumbel experiments"
echo "Sweep directory: $SWEEP_DIR"
echo "Sigmas: $SIGMAS"
echo "N_prompts: $N_PROMPTS, Max_tokens: $MAX_TOKENS"
echo "========================================================================"

###############################################################################
# Models
###############################################################################
declare -a MODELS=(
  "Qwen/Qwen3-30B-A3B"
  # "meta-llama/Llama-3.1-70B-Instruct-AWQ"   # uncomment for largest-on-1xH100 (quantized)
  # "Qwen/Qwen2.5-72B-Instruct-AWQ"           # uncomment for largest-on-1xH100 (quantized)
  "meta-llama/Llama-3.2-3B-Instruct"
  "meta-llama/Llama-3.1-8B-Instruct"
)

###############################################################################
# Helpers
###############################################################################
run_analysis() {
  local model_dir="$1"
  local model_id="$2"

  echo "Running analysis..."
  # Prefer newest gumbel_cgs_analysis_results/*; else fallback to /results
  local results_dir
  results_dir=$(ls -td "${model_dir}/gumbel_cgs_analysis_results/"* 2>/dev/null | head -1)
  if [[ -z "$results_dir" ]]; then
    results_dir="${model_dir}/results"
  fi

  if [[ ! -d "$results_dir" ]]; then
    echo "✗ No results folder found for ${model_id} (looked in ${model_dir})"
    return 1
  fi

  python adam_analyze_thresholds.py \
    --folder "$results_dir" \
    --max-thresholds "$MAX_THRESHOLDS" \
    --skip-logistic-regression

  if [[ $? -ne 0 ]]; then
    echo "✗ Analysis failed for ${model_id}"
    return 1
  fi
  echo "✓ Successfully completed analysis for ${model_id}"

  echo "Running two-step classifier analysis..."
  python adam_analyze_two_step_classifier.py \
    --folder "$results_dir" \
    --max-thresholds "$MAX_THRESHOLDS" \
    --rank-threshold 4

  if [[ $? -ne 0 ]]; then
    echo "✗ Two-step classifier analysis failed for ${model_id}"
    return 1
  fi
  echo "✓ Successfully completed two-step classifier analysis for ${model_id}"
}

###############################################################################
# Main loop
###############################################################################
for MODEL in "${MODELS[@]}"; do
  echo ""
  echo "========================================================================"
  echo "Processing model: $MODEL"
  echo "========================================================================"

  # Per-model directory & log
  MODEL_NAME=$(echo "$MODEL" | tr '/' '_')
  MODEL_DIR="${SWEEP_DIR}/${MODEL_NAME}"
  mkdir -p "$MODEL_DIR"
  LOG_FILE="${MODEL_DIR}/run.log"

  # (Optional) prefetch to local HF cache to avoid mid-run pulls
  # huggingface-cli download "$MODEL" --local-dir "${MODEL_DIR}/hf_prefetch" --resume &>/dev/null

  # Run experiment (tee logs)
  python run_gumbel_and_cgs_recording_adam_fast.py \
    --model "$MODEL" \
    --n-prompts "$N_PROMPTS" \
    --max-tokens "$MAX_TOKENS" \
    --gpu-memory-utilization "$GPU_MEM" \
    --gumbel-sigmas "$SIGMAS" \
    --sweep-dir "$MODEL_DIR" \
    --support-size "$SUPPORT_SIZE" \
    --max-model-len "$MAX_MODEL_LEN" \
    2>&1 | tee "$LOG_FILE"

  if [[ ${PIPESTATUS[0]} -eq 0 ]]; then
    echo "✓ Successfully completed experiment for $MODEL"
    run_analysis "$MODEL_DIR" "$MODEL"
  else
    echo "✗ Experiment failed for $MODEL (see log: $LOG_FILE)"
  fi
done

echo ""
echo "========================================================================"
echo "All experiments complete!"
echo "Results saved to: $SWEEP_DIR"
echo "========================================================================"
echo ""
echo "To create combined plots, run:"
echo "python plot_multi_model_comparison.py --sweep-dir $SWEEP_DIR"