playground/inference/active_cases/eval_active_cases.py

"""
Evaluate a single VLM on the 238 active (hard) cases.

Usage:
  python eval_active_cases.py \
      --model-name qwen3vl_finetuned \
      --frames-dir /mlx/users/jiashuo.fan/playground/inference/active_cases/frames_cache \
      --output-dir /mnt/bn/bohanzhainas1/jiashuo/exp/active_cases_eval

  # Override model path:
  python eval_active_cases.py --model-name internvl3_8b --model-path /some/other/path ...

Model types supported:
  qwen3vl   - Qwen3-VL (uses Qwen3VLForConditionalGeneration + qwen_vl_utils)
  qwen25vl  - Qwen2.5-VL / Qwen2-VL (uses Qwen2VLForConditionalGeneration + qwen_vl_utils)
  internvl  - InternVL (uses InternVLChatModel with trust_remote_code)
  llava     - LLaVA-OneVision (LlavaQwenForCausalLM)
  generic   - AutoModelForCausalLM with AutoProcessor / AutoTokenizer
"""

import argparse
import base64
import glob
import io
import json
import os
import re
import sys
import time
import traceback
import types
from pathlib import Path

# Stub xformers so models that optionally import it (e.g. CogVLM2) can still
# load without having xformers installed.  Replace memory_efficient_attention
# with PyTorch's scaled_dot_product_attention.
if "xformers" not in sys.modules:
    import torch as _torch
    _xops = types.ModuleType("xformers.ops")
    def _mem_eff_attn(query, key, value, attn_bias=None, scale=None, **kw):
        # xformers layout: (B, S, H, D); torch SDPA layout: (B, H, S, D)
        q = query.transpose(1, 2)
        k = key.transpose(1, 2)
        v = value.transpose(1, 2)
        out = _torch.nn.functional.scaled_dot_product_attention(
            q, k, v, attn_mask=attn_bias, scale=scale,
        )
        return out.transpose(1, 2)  # back to (B, S, H, D)
    _xops.memory_efficient_attention = _mem_eff_attn
    _xformers = types.ModuleType("xformers")
    _xformers.ops = _xops
    sys.modules["xformers"] = _xformers
    sys.modules["xformers.ops"] = _xops

from PIL import Image

# ── defaults ──────────────────────────────────────────────────────────────────

FRAMES_DIR  = Path("/mlx/users/jiashuo.fan/playground/inference/active_cases/frames_cache")
OUTPUT_DIR  = Path("/mnt/bn/bohanzhainas1/jiashuo/exp/active_cases_eval")
FRAMES_PER_VIDEO = 8     # subsample from the 16 stored frames
MAX_PIXELS       = 336 * 336
MAX_NEW_TOKENS   = 128
SAVE_INTERVAL    = 20

# ── SYSTEM prompt for non-fine-tuned models ───────────────────────────────────

BASE_SYSTEM_PROMPT = """You are an expert at analyzing pairs of TikTok videos for a "Proactive Publish" attribution task. Given two videos, you must determine whether watching Video 1 (consumption video) CAUSED or INSPIRED the user to create Video 2 (publish video).

label=1 means the videos are causally related (e.g., same meme/challenge/song, same viral format, same template).
label=0 means they are NOT causally related (they may be in the same broad category but lack direct inspiration).

You MUST respond with a JSON object and nothing else."""

BASE_USER_TEMPLATE = """Analyze these two TikTok videos:

Video 1 (consumption video - what the user watched):
{view_frames_placeholder}

Video 2 (publish video - what the user then created):
{pub_frames_placeholder}

Category: {class_name}

Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?

Respond with JSON only:
{{"reasoning": "<brief explanation>", "label": 0 or 1}}"""

# ── helpers ───────────────────────────────────────────────────────────────────

def load_sample_files(frames_dir: Path) -> list[dict]:
    files = sorted(frames_dir.glob("*.json"))
    samples = []
    for f in files:
        try:
            data = json.loads(f.read_text())
            if data.get("source") == "failed" or not data.get("messages"):
                print(f"  [SKIP] {f.stem} (no frames / failed extraction)", flush=True)
                continue
            samples.append(data)
        except Exception as e:
            print(f"  [SKIP] {f.name}: {e}", flush=True)
    return samples


def b64_to_pil(b64_str: str) -> Image.Image:
    img = Image.open(io.BytesIO(base64.b64decode(b64_str))).convert("RGB")
    w, h = img.size
    if w * h > MAX_PIXELS:
        scale = (MAX_PIXELS / (w * h)) ** 0.5
        img = img.resize((int(w * scale), int(h * scale)), Image.BILINEAR)
    return img


def subsample_frames(frames: list[str], n: int = FRAMES_PER_VIDEO) -> list[str]:
    """Pick n evenly-spaced frames from a list of base64 frames."""
    if len(frames) <= n:
        return frames
    step = len(frames) / n
    return [frames[int(i * step)] for i in range(n)]


def parse_sample_for_finetuned(sample: dict) -> tuple[list, int]:
    """
    Parse sample in training format: return (content_items, gt_label).
    content_items match the format used during training.
    """
    msgs = sample["messages"]
    user_content = msgs[0]["content"]
    gt_text = msgs[1]["content"][0]["text"] if len(msgs) > 1 else '{"label": 1}'
    gt_label = extract_label(gt_text)

    content_items = []
    for item in user_content:
        if item["type"] == "video":
            frames = subsample_frames(item["video"], FRAMES_PER_VIDEO)
            for b64 in frames:
                content_items.append({"type": "image", "image": b64_to_pil(b64)})
        elif item["type"] == "text":
            content_items.append({"type": "text", "text": item["text"]})
        elif item["type"] == "image":
            content_items.append({"type": "image", "image": b64_to_pil(item["image"])})

    return content_items, gt_label


def parse_sample_for_base(sample: dict) -> tuple[list, int]:
    """
    Parse sample for non-fine-tuned models: build a natural language prompt
    with PIL images interleaved with text.
    """
    msgs = sample["messages"]
    user_content = msgs[0]["content"]
    gt_text = msgs[1]["content"][0]["text"] if len(msgs) > 1 else '{"label": 1}'
    gt_label = extract_label(gt_text)

    class_name = sample.get("class_name", "")

    # Collect video frame lists
    video_lists = []
    for item in user_content:
        if item["type"] == "video":
            video_lists.append(item["video"])

    view_frames_b64 = subsample_frames(video_lists[0], FRAMES_PER_VIDEO) if len(video_lists) > 0 else []
    pub_frames_b64  = subsample_frames(video_lists[1], FRAMES_PER_VIDEO) if len(video_lists) > 1 else []

    view_pil = [b64_to_pil(b) for b in view_frames_b64]
    pub_pil  = [b64_to_pil(b) for b in pub_frames_b64]

    return view_pil, pub_pil, class_name, gt_label


def extract_label(text: str) -> int | None:
    try:
        stripped = text.strip()
        if "```" in stripped:
            m = re.search(r"```(?:json)?\s*([\s\S]+?)```", stripped)
            if m:
                stripped = m.group(1).strip()
        return int(json.loads(stripped)["label"])
    except Exception:
        m = re.search(r'"label"\s*:\s*([01])', text)
        return int(m.group(1)) if m else None


def compute_stats(results: list[dict]) -> dict:
    from collections import defaultdict
    label_stats = defaultdict(lambda: {"tp": 0, "fp": 0, "fn": 0, "tn": 0})
    correct = evaluated = parse_fail = 0

    for r in results:
        if "error" in r:
            continue
        if r.get("pred_label") is None:
            parse_fail += 1
            continue
        gt, pred = r.get("gt_label"), r["pred_label"]
        if gt is None:
            continue
        evaluated += 1
        if gt == pred:
            correct += 1
        for label in [0, 1]:
            if gt == label and pred == label:
                label_stats[label]["tp"] += 1
            elif gt != label and pred == label:
                label_stats[label]["fp"] += 1
            elif gt == label and pred != label:
                label_stats[label]["fn"] += 1
            else:
                label_stats[label]["tn"] += 1

    per_class = {}
    for label, s in label_stats.items():
        tp, fp, fn = s["tp"], s["fp"], s["fn"]
        prec = tp / (tp + fp) if (tp + fp) > 0 else 0.0
        rec  = tp / (tp + fn) if (tp + fn) > 0 else 0.0
        f1   = 2 * prec * rec / (prec + rec) if (prec + rec) > 0 else 0.0
        per_class[str(label)] = {
            "precision": round(prec, 4),
            "recall":    round(rec, 4),
            "f1":        round(f1, 4),
            "support":   tp + fn,
        }

    return {
        "accuracy":       round(correct / evaluated, 4) if evaluated else 0.0,
        "correct":        correct,
        "evaluated":      evaluated,
        "parse_failures": parse_fail,
        "per_class":      per_class,
    }


def save_results(out_path: Path, model_name: str, model_path: str,
                 results: list, stats: dict):
    out_path.write_text(json.dumps({
        "model_name":     model_name,
        "model_path":     model_path,
        "frames_per_video": FRAMES_PER_VIDEO,
        "max_pixels":     MAX_PIXELS,
        "total_samples":  len(results),
        **stats,
        "results": results,
    }, ensure_ascii=False, indent=2))


# ══════════════════════════════════════════════════════════════════════════════
# Model loaders
# ══════════════════════════════════════════════════════════════════════════════

def load_qwen3vl(model_path: str):
    import torch
    from transformers import Qwen3VLForConditionalGeneration, AutoProcessor

    print(f"Loading Qwen3-VL from {model_path}", flush=True)
    model = Qwen3VLForConditionalGeneration.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    return model, processor


def load_qwen25vl(model_path: str):
    import torch
    from transformers import AutoProcessor

    print(f"Loading Qwen2-VL / Qwen2.5-VL from {model_path}", flush=True)
    # Qwen2.5-VL uses Qwen2_5_VLForConditionalGeneration; Qwen2-VL uses Qwen2VLForConditionalGeneration
    try:
        from transformers import Qwen2_5_VLForConditionalGeneration
        model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            attn_implementation="flash_attention_2",
            device_map="cuda:0",
            trust_remote_code=True,
        )
        print("Loaded as Qwen2_5_VL", flush=True)
    except Exception:
        from transformers import Qwen2VLForConditionalGeneration
        model = Qwen2VLForConditionalGeneration.from_pretrained(
            model_path,
            torch_dtype=torch.bfloat16,
            attn_implementation="flash_attention_2",
            device_map="cuda:0",
            trust_remote_code=True,
        )
        print("Loaded as Qwen2VL (fallback)", flush=True)
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    return model, processor


def load_internvl(model_path: str):
    import torch
    from transformers import AutoModel, AutoTokenizer

    print(f"Loading InternVL from {model_path}", flush=True)
    model = AutoModel.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        attn_implementation="flash_attention_2",
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    return model, tokenizer


def load_llava(model_path: str):
    import torch
    from transformers import AutoProcessor, AutoModelForVision2Seq

    print(f"Loading LLaVA from {model_path}", flush=True)
    model = AutoModelForVision2Seq.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    return model, processor


def load_llama32_vision(model_path: str):
    import torch
    from transformers import MllamaForConditionalGeneration, AutoProcessor

    print(f"Loading Llama-3.2-Vision from {model_path}", flush=True)
    model = MllamaForConditionalGeneration.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path)
    return model, processor


def load_phi3_vision(model_path: str):
    import torch
    from transformers import AutoModelForCausalLM, AutoProcessor

    print(f"Loading Phi-3.5-vision from {model_path}", flush=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
        _attn_implementation="flash_attention_2",
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True, num_crops=4)
    return model, processor


def load_minicpm_v(model_path: str):
    import torch
    from transformers import AutoModel, AutoTokenizer

    print(f"Loading MiniCPM-V from {model_path}", flush=True)
    model = AutoModel.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    return model, tokenizer


def load_pixtral(model_path: str):
    import torch
    from transformers import LlavaForConditionalGeneration, AutoProcessor

    print(f"Loading Pixtral from {model_path}", flush=True)
    model = LlavaForConditionalGeneration.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path)
    return model, processor


def load_janus(model_path: str):
    import torch
    from transformers import AutoModelForCausalLM, AutoProcessor

    print(f"Loading Janus-Pro from {model_path}", flush=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    return model, processor


def load_cogvlm2(model_path: str):
    import torch
    from transformers import AutoTokenizer, AutoModelForCausalLM
    print(f"Loading CogVLM2 from {model_path}", flush=True)
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    return model, tokenizer


def run_cogvlm2(model, tokenizer, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for CogVLM2 (requires build_conversation_input_ids)."""
    import torch
    from PIL import Image

    # CogVLM2 supports only a single image; create a 4x2 grid:
    #   top row: 4 frames from video 1, bottom row: 4 frames from video 2
    n = min(4, len(view_pil))
    m = min(4, len(pub_pil))
    frames = view_pil[:n] + pub_pil[:m]
    cell_w, cell_h = 224, 224
    cols = 4
    rows = (len(frames) + cols - 1) // cols
    grid = Image.new("RGB", (cols * cell_w, rows * cell_h), (0, 0, 0))
    for idx, fr in enumerate(frames):
        fr_r = fr.resize((cell_w, cell_h))
        grid.paste(fr_r, ((idx % cols) * cell_w, (idx // cols) * cell_h))

    query = (
        f"The image shows a grid of video frames: "
        f"top row has {n} frames from Video 1 (consumption), "
        f"bottom row has {m} frames from Video 2 (publish). "
        f"Category: {class_name}. "
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2? "
        "label=1: causally related, label=0: not causally related. "
        'JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    input_by_model = model.build_conversation_input_ids(
        tokenizer,
        query=query,
        history=[],
        images=[grid],
        template_version="chat",
    )
    device = next(model.parameters()).device
    inputs = {
        "input_ids":      input_by_model["input_ids"].unsqueeze(0).to(device),
        "token_type_ids": input_by_model["token_type_ids"].unsqueeze(0).to(device),
        "attention_mask": input_by_model["attention_mask"].unsqueeze(0).to(device),
        "images":         [[img.to(device).to(torch.bfloat16)
                            for img in input_by_model["images"]]],
    }
    # Patch methods removed from transformers >= 4.46 that CogVLM2 relies on.
    if not hasattr(model, "_extract_past_from_model_output"):
        def _extract_past(model_output, standardize_cache_format=False):
            return getattr(model_output, "past_key_values", None)
        model._extract_past_from_model_output = _extract_past

    # Patch llm_forward to handle ((None,None),...) past_key_values from newer transformers
    _orig_llm_forward = model.model.llm_forward
    def _patched_llm_forward(self_or_first, *args, **kwargs):
        # Detect if called as bound method (self is model.model) or unbound
        if callable(_orig_llm_forward):
            # get past_key_values from kwargs or args
            pkv = kwargs.get("past_key_values", None)
            if pkv is not None and hasattr(pkv, "__len__"):
                # If all layers are None, treat as None
                if all((layer is None or (hasattr(layer, "__len__") and all(t is None for t in layer)))
                       for layer in pkv):
                    kwargs["past_key_values"] = None
        return _orig_llm_forward(*args, **kwargs) if args else _orig_llm_forward(**kwargs)
    # simpler: just patch at the module level
    import types as _types
    def _safe_llm_forward(self, *args, **kwargs):
        pkv = kwargs.get("past_key_values", None)
        if pkv is not None and hasattr(pkv, "__len__"):
            if all((layer is None or (hasattr(layer, "__len__") and all(t is None for t in layer)))
                   for layer in pkv):
                kwargs["past_key_values"] = None
        return _orig_llm_forward(*args, **kwargs)
    model.model.llm_forward = _types.MethodType(_safe_llm_forward, model.model)

    gen_kwargs = {
        "max_new_tokens": MAX_NEW_TOKENS,
        "pad_token_id":   tokenizer.eos_token_id,
        "do_sample":      False,
    }
    with torch.no_grad():
        outputs = model.generate(**inputs, **gen_kwargs)
        outputs = outputs[:, inputs["input_ids"].shape[1]:]
    return tokenizer.decode(outputs[0], skip_special_tokens=True)


def load_molmo(model_path: str):
    import sys, torch
    from transformers import AutoModelForCausalLM, AutoProcessor

    # Molmo's image_preprocessing_molmo.py has a conditional `import tensorflow` that
    # causes check_imports to fail if tensorflow is not installed.  Stub it out.
    if "tensorflow" not in sys.modules:
        # Create a permissive tensorflow stub: any attribute access returns a
        # no-op callable/class so Molmo's check_imports and processor don't crash.
        class _TFStub(types.ModuleType):
            def __getattr__(self, name):
                if name.startswith("_"):
                    raise AttributeError(name)
                # Return a dummy class / callable for unknown attrs
                dummy = type(name, (), {"__call__": lambda self, *a, **kw: False})()
                setattr(self, name, dummy)
                return dummy
        tf_stub = _TFStub("tensorflow")
        tf_stub.Tensor = type("Tensor", (), {})
        tf_stub.Variable = type("Variable", (), {})
        tf_stub.is_tensor = lambda x: False
        tf_stub.string = str
        tf_stub.float32 = "float32"
        tf_stub.int32 = "int32"
        _keras = _TFStub("tensorflow.keras")
        _keras_backend = _TFStub("tensorflow.keras.backend")
        _keras_backend.image_data_format = lambda: "channels_last"
        _keras.backend = _keras_backend
        tf_stub.keras = _keras
        tf_stub.io = _TFStub("tensorflow.io")
        sys.modules["tensorflow"] = tf_stub
        sys.modules["tensorflow.io"] = tf_stub.io
        sys.modules["tensorflow.keras"] = _keras
        sys.modules["tensorflow.keras.backend"] = _keras_backend

    print(f"Loading Molmo from {model_path}", flush=True)
    processor = AutoProcessor.from_pretrained(
        model_path, trust_remote_code=True,
        torch_dtype=torch.bfloat16,
    )
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    return model, processor


def load_moondream(model_path: str):
    import shutil, torch
    from pathlib import Path
    from transformers import AutoModelForCausalLM, AutoTokenizer

    # Pre-populate HF modules cache with ALL .py files from the local model dir
    # so that transformers' dynamic_module_utils can resolve all relative imports.
    model_name = Path(model_path).name
    cache_dir = Path.home() / ".cache" / "huggingface" / "modules" / "transformers_modules" / model_name
    if cache_dir.exists():
        shutil.rmtree(cache_dir)
    cache_dir.mkdir(parents=True, exist_ok=True)
    for py_file in Path(model_path).glob("*.py"):
        shutil.copy2(py_file, cache_dir / py_file.name)
    print(f"Pre-populated cache with {len(list(cache_dir.glob('*.py')))} .py files", flush=True)

    print(f"Loading Moondream from {model_path}", flush=True)
    model = AutoModelForCausalLM.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map="cuda:0",
        trust_remote_code=True,
    )
    model.eval()
    tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
    return model, tokenizer


def load_generic(model_path: str):
    """Generic loader: try AutoModelForCausalLM, fallback to AutoModelForVision2Seq, then AutoModel."""
    import torch
    from transformers import AutoModelForCausalLM, AutoModelForVision2Seq, AutoModel, AutoProcessor, AutoTokenizer

    print(f"Loading generic model from {model_path}", flush=True)
    try:
        processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    except Exception:
        processor = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

    load_kwargs = dict(torch_dtype=torch.bfloat16, device_map="cuda:0", trust_remote_code=True)
    model = None
    for cls in (AutoModelForCausalLM, AutoModelForVision2Seq, AutoModel):
        try:
            model = cls.from_pretrained(model_path, **load_kwargs)
            print(f"Loaded with {cls.__name__}", flush=True)
            break
        except Exception as e:
            print(f"  {cls.__name__} failed: {e}", flush=True)
    if model is None:
        raise RuntimeError(f"Could not load model from {model_path} with any loader")
    model.eval()
    return model, processor


# ══════════════════════════════════════════════════════════════════════════════
# Inference functions
# ══════════════════════════════════════════════════════════════════════════════

def run_qwenvl_finetuned(model, processor, content_items: list) -> str:
    """Inference for fine-tuned Qwen3-VL / Qwen2.5-VL (uses training message format)."""
    import torch
    from qwen_vl_utils import process_vision_info

    messages = [{"role": "user", "content": content_items}]
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    image_inputs, _ = process_vision_info(messages)

    inputs = processor(
        text=[text],
        images=image_inputs,
        return_tensors="pt",
    ).to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
            pad_token_id=processor.tokenizer.eos_token_id,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_qwenvl_base(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for base/instruct Qwen3-VL / Qwen2.5-VL (natural language prompt)."""
    import torch
    from qwen_vl_utils import process_vision_info

    # Build content: system + interleaved images + text
    content = []

    content.append({"type": "text", "text": "Video 1 (consumption video):"})
    for img in view_pil:
        content.append({"type": "image", "image": img})

    content.append({"type": "text", "text": "\nVideo 2 (publish video):"})
    for img in pub_pil:
        content.append({"type": "image", "image": img})

    content.append({
        "type": "text",
        "text": (
            f"\nCategory: {class_name}\n\n"
            "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
            "label=1: causally related (same meme/challenge/song/template)\n"
            "label=0: not causally related\n\n"
            'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
        ),
    })

    messages = [
        {"role": "system", "content": BASE_SYSTEM_PROMPT},
        {"role": "user",   "content": content},
    ]
    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    image_inputs, _ = process_vision_info(messages)

    inputs = processor(
        text=[text],
        images=image_inputs,
        return_tensors="pt",
    ).to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
            pad_token_id=processor.tokenizer.eos_token_id,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_internvl(model, tokenizer, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for InternVL models."""
    import torch
    import torchvision.transforms as T
    from torchvision.transforms.functional import InterpolationMode

    IMAGENET_MEAN = (0.485, 0.456, 0.406)
    IMAGENET_STD  = (0.229, 0.224, 0.225)

    # Use 448 with downsample; each image → 1 tile → 256 tokens after pixel shuffle
    def build_transform(input_size=448):
        return T.Compose([
            T.Lambda(lambda img: img.convert("RGB")),
            T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
            T.ToTensor(),
            T.Normalize(mean=IMAGENET_MEAN, std=IMAGENET_STD),
        ])

    transform = build_transform(448)
    all_images = view_pil + pub_pil
    pixel_values = torch.stack([transform(img) for img in all_images]).to(torch.bfloat16).cuda()

    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    # InternVL uses plain <image>\n tokens — one per image
    view_img_tokens = "<image>\n" * n_view
    pub_img_tokens  = "<image>\n" * n_pub

    question = (
        f"Video 1 (consumption video) - {n_view} frames:\n{view_img_tokens}"
        f"Video 2 (publish video) - {n_pub} frames:\n{pub_img_tokens}"
        f"Category: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    # num_patches_list: one tile per image
    num_patches_list = [1] * len(all_images)
    generation_config = dict(max_new_tokens=MAX_NEW_TOKENS, do_sample=False)

    response = model.chat(
        tokenizer,
        pixel_values,
        question,
        generation_config,
        num_patches_list=num_patches_list,
        history=None,
        return_history=False,
    )
    return response


def run_llava(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for LLaVA-OneVision."""
    import torch

    # Build conversation with image tokens
    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    view_img_str = "\n".join(["<image>"] * n_view)
    pub_img_str  = "\n".join(["<image>"] * n_pub)

    text_prompt = (
        f"Video 1 (consumption video):\n{view_img_str}\n\n"
        f"Video 2 (publish video):\n{pub_img_str}\n\n"
        f"Category: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    conversation = [
        {"role": "system", "content": BASE_SYSTEM_PROMPT},
        {"role": "user",   "content": text_prompt},
    ]

    all_images = view_pil + pub_pil
    prompt = processor.apply_chat_template(
        conversation, tokenize=False, add_generation_prompt=True
    )
    inputs = processor(
        images=all_images,
        text=prompt,
        return_tensors="pt",
    ).to("cuda:0")
    inputs = {k: v.to(torch.bfloat16) if v.dtype == torch.float32 else v
              for k, v in inputs.items()}

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_llama32_vision(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Llama-3.2-Vision."""
    import torch

    all_images = view_pil + pub_pil
    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    view_img_str = "".join([f"<|image|>" for _ in view_pil])
    pub_img_str  = "".join([f"<|image|>" for _ in pub_pil])

    messages = [
        {"role": "user", "content": [
            {"type": "text", "text": (
                f"Video 1 (consumption video) - {n_view} frames:\n"
            )},
            *[{"type": "image"} for _ in view_pil],
            {"type": "text", "text": (
                f"\nVideo 2 (publish video) - {n_pub} frames:\n"
            )},
            *[{"type": "image"} for _ in pub_pil],
            {"type": "text", "text": (
                f"\nCategory: {class_name}\n\n"
                "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
                "label=1: causally related, label=0: not causally related\n\n"
                'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
            )},
        ]},
    ]

    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(
        images=all_images,
        text=text,
        return_tensors="pt",
    ).to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_phi3_vision(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Phi-3.5-vision."""
    import torch

    # Phi-3.5 uses seen_tokens / get_usable_length on DynamicCache; patch if missing
    try:
        from transformers.cache_utils import DynamicCache
        if not hasattr(DynamicCache, "seen_tokens"):
            DynamicCache.seen_tokens = property(lambda self: self.get_seq_length())
        if not hasattr(DynamicCache, "get_usable_length"):
            def _get_usable_length(self, new_seq_length, layer_idx=0):
                return self.get_seq_length(layer_idx)
            DynamicCache.get_usable_length = _get_usable_length
    except Exception:
        pass

    all_images = view_pil + pub_pil
    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    view_img_tags = "".join([f"<|image_{i+1}|>\n" for i in range(n_view)])
    pub_img_tags  = "".join([f"<|image_{n_view+i+1}|>\n" for i in range(n_pub)])

    messages = [
        {"role": "user", "content": (
            f"Video 1 (consumption video):\n{view_img_tags}"
            f"Video 2 (publish video):\n{pub_img_tags}"
            f"Category: {class_name}\n\n"
            "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
            "label=1: causally related, label=0: not causally related\n\n"
            'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
        )},
    ]

    prompt = processor.tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    inputs = processor(prompt, all_images, return_tensors="pt").to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
            eos_token_id=processor.tokenizer.eos_token_id,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.tokenizer.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_minicpm_v(model, tokenizer, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for MiniCPM-V."""
    import torch

    all_images = view_pil + pub_pil

    question = (
        f"Video 1 (consumption video) - {len(view_pil)} frames (shown above)\n"
        f"Video 2 (publish video) - {len(pub_pil)} frames (shown above)\n\n"
        f"Category: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    msgs = [{"role": "user", "content": all_images + [question]}]

    res = model.chat(
        image=None,
        msgs=msgs,
        tokenizer=tokenizer,
        sampling=False,
        max_new_tokens=MAX_NEW_TOKENS,
    )
    return res


def run_pixtral(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Pixtral-12B."""
    import torch

    all_images = view_pil + pub_pil
    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    content = []
    content.append({"type": "text", "text": f"Video 1 (consumption video) - {n_view} frames:"})
    for _ in view_pil:
        content.append({"type": "image"})
    content.append({"type": "text", "text": f"\nVideo 2 (publish video) - {n_pub} frames:"})
    for _ in pub_pil:
        content.append({"type": "image"})
    content.append({"type": "text", "text": (
        f"\nCategory: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )})

    messages = [{"role": "user", "content": content}]

    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(
        images=all_images,
        text=text,
        return_tensors="pt",
    ).to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )

    prompt_len = inputs["input_ids"].shape[1]
    return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)


def run_janus(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Janus-Pro (DeepSeek multi-modal understanding)."""
    import torch

    all_images = view_pil + pub_pil
    n_view, n_pub = len(view_pil), len(pub_pil)

    # Janus uses image tokens in conversation format
    img_tags_view = "<image_placeholder>" * n_view
    img_tags_pub  = "<image_placeholder>" * n_pub

    conversation = [
        {"role": "User", "content": (
            f"Video 1 (consumption video) - {n_view} frames:\n{img_tags_view}\n"
            f"Video 2 (publish video) - {n_pub} frames:\n{img_tags_pub}\n\n"
            f"Category: {class_name}\n\n"
            "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
            "label=1: causally related, label=0: not causally related\n\n"
            'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
        )},
        {"role": "Assistant", "content": ""},
    ]

    prepare = processor(
        conversations=conversation,
        images=all_images,
        force_batchify=True,
    ).to("cuda:0")

    inputs_embeds = model.prepare_inputs_embeds(**prepare)
    with torch.no_grad():
        output_ids = model.language_model.generate(
            inputs_embeds=inputs_embeds,
            attention_mask=prepare.attention_mask,
            pad_token_id=processor.tokenizer.eos_token_id,
            bos_token_id=processor.tokenizer.bos_token_id,
            eos_token_id=processor.tokenizer.eos_token_id,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )
    return processor.tokenizer.decode(output_ids[0].cpu(), skip_special_tokens=True)


def run_molmo(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Molmo (AllenAI)."""
    import torch

    all_images = view_pil + pub_pil

    prompt = (
        f"Video 1 (consumption video) - {len(view_pil)} frames and "
        f"Video 2 (publish video) - {len(pub_pil)} frames are shown above.\n"
        f"Category: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    inputs = processor.process(
        images=all_images,
        text=prompt,
    )
    inputs = {k: v.to("cuda:0").unsqueeze(0) if hasattr(v, "to") else v
              for k, v in inputs.items()}

    from transformers import GenerationConfig
    gen_cfg = getattr(model.config, "generation_config", None) or GenerationConfig(
        max_new_tokens=MAX_NEW_TOKENS, do_sample=False,
    )
    with torch.no_grad():
        output = model.generate_from_batch(
            inputs,
            generation_config=gen_cfg,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )
    generated = output[0, inputs["input_ids"].size(1):]
    return processor.tokenizer.decode(generated, skip_special_tokens=True)


def run_moondream(model, tokenizer, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for Moondream2."""
    # Moondream encodes each image independently, then does text generation
    question = (
        f"These are frames from two TikTok videos. "
        f"Video 1 ({len(view_pil)} frames) then Video 2 ({len(pub_pil)} frames). "
        f"Category: {class_name}. "
        "Did Video 1 CAUSE or INSPIRE the creation of Video 2? "
        "label=1: yes, label=0: no. "
        'JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    all_images = view_pil + pub_pil
    enc_images = [model.encode_image(img) for img in all_images]

    # Use first encoded image as primary, append others in question context
    answer = model.answer_question(
        enc_images[0],
        question,
        tokenizer,
        max_new_tokens=MAX_NEW_TOKENS,
    )
    return answer


def run_generic(model, processor, view_pil: list, pub_pil: list, class_name: str) -> str:
    """Inference for generic AutoModel (best-effort)."""
    import torch

    # Patch config.num_hidden_layers if missing (e.g. ChatGLM uses num_layers)
    cfg = getattr(model, "config", None)
    if cfg is not None and not hasattr(cfg, "num_hidden_layers"):
        for alt in ("num_layers", "n_layer", "n_layers"):
            if hasattr(cfg, alt):
                cfg.num_hidden_layers = getattr(cfg, alt)
                break

    all_images = view_pil + pub_pil
    n_view = len(view_pil)
    n_pub  = len(pub_pil)

    text_question = (
        f"Category: {class_name}\n\n"
        "Did watching Video 1 CAUSE or INSPIRE the creation of Video 2?\n"
        "label=1: causally related, label=0: not causally related\n\n"
        'Respond with JSON only: {"reasoning": "...", "label": 0 or 1}'
    )

    # Build messages using dict content (works for Idefics3 and most VLMs)
    # String content with <image> tokens gets stripped by some chat templates.
    user_content = (
        [{"type": "text", "text": "Video 1 (consumption video):"}]
        + [{"type": "image"} for _ in view_pil]
        + [{"type": "text", "text": "\nVideo 2 (publish video):"}]
        + [{"type": "image"} for _ in pub_pil]
        + [{"type": "text", "text": f"\n{text_question}"}]
    )
    messages = [
        {"role": "system", "content": BASE_SYSTEM_PROMPT},
        {"role": "user", "content": user_content},
    ]

    # Try dict-content template first; fall back to string-content template
    prompt = None
    if hasattr(processor, "apply_chat_template"):
        try:
            prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
            # Verify image tokens are present
            img_token = getattr(processor, "image_token", "<image>")
            if img_token not in prompt and "<image" not in prompt:
                raise ValueError("no image tokens in template output")
        except Exception:
            # Fall back: string content with <image> placeholders
            view_img_str = "\n".join(["<image>"] * n_view)
            pub_img_str  = "\n".join(["<image>"] * n_pub)
            messages_str = [
                {"role": "system", "content": BASE_SYSTEM_PROMPT},
                {"role": "user", "content": (
                    f"Video 1 (consumption video):\n{view_img_str}\n\n"
                    f"Video 2 (publish video):\n{pub_img_str}\n\n{text_question}"
                )},
            ]
            try:
                prompt = processor.apply_chat_template(messages_str, tokenize=False, add_generation_prompt=True)
            except Exception:
                prompt = messages_str[-1]["content"]
    if prompt is None:
        view_img_str = "\n".join(["<image>"] * n_view)
        pub_img_str  = "\n".join(["<image>"] * n_pub)
        prompt = (f"Video 1:\n{view_img_str}\nVideo 2:\n{pub_img_str}\n{text_question}")

    try:
        if hasattr(processor, "image_processor") or hasattr(processor, "feature_extractor"):
            # Idefics3/SmolVLM requires images as List[List[PIL.Image]] (batch of samples)
            proc_cls = type(processor).__name__
            if "Idefics" in proc_cls or "SmolVLM" in proc_cls or hasattr(processor, "image_seq_len"):
                images_arg = [all_images]
            else:
                images_arg = all_images
            inputs = processor(
                images=images_arg,
                text=prompt,
                return_tensors="pt",
            ).to("cuda:0")
        else:
            inputs = processor(prompt, return_tensors="pt").to("cuda:0")
    except Exception:
        inputs = processor(prompt, return_tensors="pt").to("cuda:0")

    with torch.no_grad():
        output_ids = model.generate(
            **inputs,
            max_new_tokens=MAX_NEW_TOKENS,
            do_sample=False,
        )

    if hasattr(processor, "decode"):
        prompt_len = inputs["input_ids"].shape[1]
        return processor.decode(output_ids[0, prompt_len:], skip_special_tokens=True)
    else:
        return processor.batch_decode(output_ids, skip_special_tokens=True)[0]


# ══════════════════════════════════════════════════════════════════════════════
# Main
# ══════════════════════════════════════════════════════════════════════════════

def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--model-name",   required=True,
                        help="Model name from models.py registry (e.g. qwen3vl_finetuned)")
    parser.add_argument("--model-path",   default=None,
                        help="Override model path (otherwise uses registry)")
    parser.add_argument("--model-type",   default=None,
                        help="Override model type: qwen3vl|qwen25vl|internvl|llava|generic")
    parser.add_argument("--finetuned",    action="store_true",
                        help="Use training-format prompt (for fine-tuned models)")
    parser.add_argument("--frames-dir",      default=str(FRAMES_DIR))
    parser.add_argument("--output-dir",      default=str(OUTPUT_DIR))
    parser.add_argument("--gpu-id",          type=int, default=0)
    parser.add_argument("--frames-per-video", type=int, default=None,
                        help="Override FRAMES_PER_VIDEO (default: 8)")
    args = parser.parse_args()

    if args.frames_per_video is not None:
        global FRAMES_PER_VIDEO
        FRAMES_PER_VIDEO = args.frames_per_video

    os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id)

    # Resolve model config from registry
    sys.path.insert(0, str(Path(__file__).parent))
    from models import MODELS_BY_NAME

    registry = MODELS_BY_NAME.get(args.model_name, {})
    model_path = args.model_path or registry.get("model_path", args.model_name)
    model_type = args.model_type or registry.get("model_type", "generic")

    # Fine-tuned flag: auto-detect if model name is qwen3vl_finetuned, else use --finetuned flag
    is_finetuned = args.finetuned or (args.model_name == "qwen3vl_finetuned")

    print(f"Model:      {args.model_name}", flush=True)
    print(f"Model path: {model_path}", flush=True)
    print(f"Model type: {model_type}", flush=True)
    print(f"Fine-tuned: {is_finetuned}", flush=True)

    # Output
    out_dir  = Path(args.output_dir)
    out_dir.mkdir(parents=True, exist_ok=True)
    out_path = out_dir / f"{args.model_name}.json"

    # Load existing results (resume)
    done_keys: set[str] = set()
    results: list[dict] = []
    if out_path.exists():
        try:
            saved = json.loads(out_path.read_text())
            results = saved.get("results", [])
            done_keys = {r["key"] for r in results if "key" in r}
            print(f"Resuming: {len(done_keys)} already done", flush=True)
        except Exception:
            pass

    # Load samples
    frames_dir = Path(args.frames_dir)
    samples = load_sample_files(frames_dir)
    print(f"Loaded {len(samples)} samples from {frames_dir}", flush=True)

    pending = [s for s in samples
               if f"{s['view_gid']}_{s['pub_gid']}" not in done_keys]
    print(f"Pending: {len(pending)}", flush=True)

    if not pending:
        print("Nothing to evaluate!", flush=True)
        return

    # Load model
    loader_map = {
        "qwen3vl":         load_qwen3vl,
        "qwen25vl":        load_qwen25vl,
        "internvl":        load_internvl,
        "llava":           load_llava,
        "llama32_vision":  load_llama32_vision,
        "phi3_vision":     load_phi3_vision,
        "minicpm_v":       load_minicpm_v,
        "pixtral":         load_pixtral,
        "janus":           load_janus,
        "molmo":           load_molmo,
        "moondream":       load_moondream,
        "cogvlm2":         load_cogvlm2,
        "generic":         load_generic,
    }
    loader = loader_map.get(model_type, load_generic)
    model, processor = loader(model_path)

    # Inference loop
    t0 = time.time()
    for i, sample in enumerate(pending):
        key = f"{sample['view_gid']}_{sample['pub_gid']}"
        result = {
            "key":        key,
            "view_gid":   sample["view_gid"],
            "pub_gid":    sample["pub_gid"],
            "class_name": sample.get("class_name", ""),
        }

        try:
            if is_finetuned and model_type in ("qwen3vl", "qwen25vl"):
                content_items, gt_label = parse_sample_for_finetuned(sample)
                pred_text = run_qwenvl_finetuned(model, processor, content_items)
            else:
                view_pil, pub_pil, class_name, gt_label = parse_sample_for_base(sample)

                if model_type in ("qwen3vl", "qwen25vl"):
                    pred_text = run_qwenvl_base(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "internvl":
                    pred_text = run_internvl(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "llava":
                    pred_text = run_llava(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "llama32_vision":
                    pred_text = run_llama32_vision(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "phi3_vision":
                    pred_text = run_phi3_vision(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "minicpm_v":
                    pred_text = run_minicpm_v(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "pixtral":
                    pred_text = run_pixtral(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "janus":
                    pred_text = run_janus(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "molmo":
                    pred_text = run_molmo(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "moondream":
                    pred_text = run_moondream(model, processor, view_pil, pub_pil, class_name)
                elif model_type == "cogvlm2":
                    pred_text = run_cogvlm2(model, processor, view_pil, pub_pil, class_name)
                else:
                    pred_text = run_generic(model, processor, view_pil, pub_pil, class_name)

            pred_label = extract_label(pred_text)
            result.update({
                "gt_label":   gt_label,
                "pred_label": pred_label,
                "match":      (pred_label == gt_label) if (pred_label is not None and gt_label is not None) else None,
                "prediction": pred_text,
            })
        except Exception as e:
            result["error"] = str(e)
            result["traceback"] = traceback.format_exc()[:500]
            print(f"  ERROR on {key}: {e}", flush=True)

        results.append(result)
        done_keys.add(key)

        # Progress
        elapsed = time.time() - t0
        speed = (i + 1) / elapsed
        total_done = len(done_keys)
        stats = compute_stats(results)
        print(
            f"[{total_done}/{len(samples)}] {key} | "
            f"acc={stats['accuracy']:.3f} "
            f"(correct={stats['correct']}/{stats['evaluated']}) "
            f"| {speed:.2f} samp/s",
            flush=True,
        )

        # Save periodically
        if (i + 1) % SAVE_INTERVAL == 0:
            stats = compute_stats(results)
            save_results(out_path, args.model_name, model_path, results, stats)

    # Final save
    stats = compute_stats(results)
    save_results(out_path, args.model_name, model_path, results, stats)

    elapsed = time.time() - t0
    print(f"\n{'='*60}", flush=True)
    print(f"DONE  model={args.model_name}", flush=True)
    print(f"  accuracy:  {stats['accuracy']:.4f} ({stats['correct']}/{stats['evaluated']})", flush=True)
    print(f"  per-class: {json.dumps(stats['per_class'], indent=4)}", flush=True)
    print(f"  parse_failures: {stats['parse_failures']}", flush=True)
    print(f"  time: {elapsed:.1f}s  ({len(results)/elapsed:.2f} samp/s)", flush=True)
    print(f"  saved -> {out_path}", flush=True)


if __name__ == "__main__":
    main()