playground/inference/active_cases/run_parallel_models.sh

#!/bin/bash
# Run 20-model active-cases eval in parallel across 8 GPUs.
# Models are assigned to GPUs in round-robin batches of 8.
# Each model runs on a dedicated GPU → full utilization.
#
# Usage:
#   bash run_parallel_models.sh [--skip-extract] [--models m1,m2,...]

set -uo pipefail

SCRIPT_DIR=/mlx/users/jiashuo.fan/playground/inference/active_cases
PYTHON_ABBIE=/mlx/users/jiashuo.fan/miniconda3/envs/abbie/bin/python3
LOG_DIR=$SCRIPT_DIR/logs
FRAMES_DIR=$SCRIPT_DIR/frames_cache
OUTPUT_DIR=/mnt/bn/bohanzhainas1/jiashuo/exp/active_cases_eval

mkdir -p "$LOG_DIR" "$FRAMES_DIR" "$OUTPUT_DIR"

SKIP_EXTRACT=false
ONLY_MODELS=""
for arg in "$@"; do
    case $arg in
        --skip-extract)   SKIP_EXTRACT=true   ;;
        --models=*)       ONLY_MODELS="${arg#*=}" ;;
    esac
done

# ── Step 1: Frame extraction ─────────────────────────────────────────────────
if [ "$SKIP_EXTRACT" = false ]; then
    echo "========================================================"
    echo "STEP 1: Extracting frames (if not already done)"
    echo "========================================================"
    n_frames=$(ls "$FRAMES_DIR"/*.json 2>/dev/null | wc -l)
    if [ "$n_frames" -ge 200 ]; then
        echo "[SKIP] Already have $n_frames frame JSON files."
    else
        LOG_EXTRACT=$LOG_DIR/extract_frames.log
        $PYTHON_ABBIE -u $SCRIPT_DIR/extract_frames.py \
            --output-dir "$FRAMES_DIR" 2>&1 | tee "$LOG_EXTRACT"
        echo "Frame extraction done."
    fi
fi

# ── Model list: name:type:flags ──────────────────────────────────────────────
ALL_MODELS=(
    "qwen3vl_finetuned:qwen3vl:--finetuned"
    "qwen3vl_base:qwen3vl:"
    "qwen3vl_instruct:qwen3vl:"
    "qwen25vl_7b:qwen25vl:"
    "internvl3_8b:internvl:"
    "internvl35:internvl:"
    "llava_ov:llava:"
    "llama32_11b_vision:llama32_vision:"
    "phi35_vision:phi3_vision:"
    "minicpm_v25:minicpm_v:"
    "qwen2vl_7b:qwen25vl:"
    "pixtral_12b:pixtral:"
    "smolvlm:generic:"
    "glm4v_9b:generic:"
    "idefics3_8b:generic:"
    "phi4_multimodal:generic:"
    "cogvlm2_19b:generic:"
    "deepseek_vl2_small:generic:"
    "videollama2_7b:generic:"
    "gemma3_4b:generic:"
    "janus_pro_7b:janus:"
    "molmo_7b:molmo:"
    "moondream2:moondream:"
)

# Filter to requested models if --models= was specified
if [ -n "$ONLY_MODELS" ]; then
    IFS=',' read -ra REQUESTED <<< "$ONLY_MODELS"
    FILTERED=()
    for entry in "${ALL_MODELS[@]}"; do
        model_name="${entry%%:*}"
        for req in "${REQUESTED[@]}"; do
            [ "$model_name" = "$req" ] && FILTERED+=("$entry")
        done
    done
    ALL_MODELS=("${FILTERED[@]:-}")
fi

# Filter out already-completed models
PENDING=()
for entry in "${ALL_MODELS[@]}"; do
    model_name="${entry%%:*}"
    OUT_FILE="$OUTPUT_DIR/${model_name}.json"
    if [ -f "$OUT_FILE" ]; then
        N_RESULTS=$($PYTHON_ABBIE -c "
import json
d = json.load(open('$OUT_FILE'))
print(d.get('evaluated', 0))
" 2>/dev/null || echo "0")
        if [ "$N_RESULTS" -ge 200 ]; then
            echo "[DONE] $model_name ($N_RESULTS evaluated) — skipping"
            continue
        fi
    fi
    PENDING+=("$entry")
done

echo ""
echo "========================================================"
echo "STEP 2: Running ${#PENDING[@]} models on 8 GPUs"
echo "========================================================"
echo ""

if [ ${#PENDING[@]} -eq 0 ]; then
    echo "All models already evaluated!"
else
    # Process in batches of 8 (one per GPU)
    i=0
    while [ $i -lt ${#PENDING[@]} ]; do
        BATCH_PIDS=()
        BATCH_NAMES=()
        GPU_ID=0
        echo "--- Batch starting at index $i ---"

        for (( j=i; j<i+8 && j<${#PENDING[@]}; j++ )); do
            entry="${PENDING[$j]}"
            model_name="${entry%%:*}"
            rest="${entry#*:}"
            model_type="${rest%%:*}"
            extra_flags="${rest#*:}"

            # Check model path exists
            MODEL_PATH=$($PYTHON_ABBIE -c "
import sys; sys.path.insert(0,'$SCRIPT_DIR')
from models import MODELS_BY_NAME
m = MODELS_BY_NAME.get('$model_name', {})
print(m.get('model_path', ''))
" 2>/dev/null)
            if [ -z "$MODEL_PATH" ] || [ ! -d "$MODEL_PATH" ]; then
                echo "[SKIP] $model_name — model path not found: $MODEL_PATH"
                GPU_ID=$((GPU_ID + 1))
                continue
            fi

            LOG_FILE="$LOG_DIR/eval_${model_name}.log"
            CMD="CUDA_VISIBLE_DEVICES=$GPU_ID $PYTHON_ABBIE -u $SCRIPT_DIR/eval_active_cases.py \
                --model-name  $model_name \
                --model-type  $model_type \
                --frames-dir  $FRAMES_DIR \
                --output-dir  $OUTPUT_DIR \
                --gpu-id      $GPU_ID"
            [ -n "$extra_flags" ] && CMD="$CMD $extra_flags"

            echo "  GPU$GPU_ID ← $model_name  (log: $LOG_FILE)"
            eval "$CMD" >> "$LOG_FILE" 2>&1 &
            BATCH_PIDS+=($!)
            BATCH_NAMES+=("$model_name")
            GPU_ID=$((GPU_ID + 1))
        done

        # Wait for all in this batch to finish
        echo "Waiting for batch: ${BATCH_NAMES[*]}"
        for pid in "${BATCH_PIDS[@]}"; do
            wait "$pid" || true
        done
        echo "Batch done: $(date)"

        i=$((i + 8))
    done
fi

# ── Step 3: Summary table ────────────────────────────────────────────────────
echo ""
echo "========================================================"
echo "FINAL SUMMARY"
echo "========================================================"
$PYTHON_ABBIE -u $SCRIPT_DIR/compare_results.py \
    --results-dir "$OUTPUT_DIR" 2>/dev/null || echo "(no results yet)"

echo ""
echo "Results at: $OUTPUT_DIR"
echo "Logs at:    $LOG_DIR"