eval/vqa/run_single_model_native.py

import argparse
import json
import math
import os
import random
import re
import sys
from pathlib import Path
from typing import Optional, Tuple

import torch
from PIL import Image
from transformers import AutoTokenizer

from internvl.conversation import get_conv_template
from internvl.conversation import register_conv_template
from internvl.conversation import Conversation
from internvl.conversation import SeparatorStyle
from internvl.model.internvl_chat import InternVLChatModel
from internvl.model.internvl_chat.configuration_internvl_chat import InternVLChatConfig
from internvl.train.dataset import build_transform, dynamic_preprocess

from evaluate_vqa import VQADataset, ds_collections
from textvqa_eval import TextVQAAccuracyEvaluator


BASE_PROMPT = "Answer the question using a single word or phrase."
VIZWIZ_PROMPT = "When the provided information is insufficient, respond with 'Unanswerable'. "
INFOGRAPHICSVQA_PROMPT = "Answer the question using a single word or phrase."
AI2D_PROMPT = ""
HIDDEN_REASONING_INSTRUCTION = (
    "Think through the relevant visual evidence and any text in the image step by step internally before answering."
)
EXPLICIT_REASONING_INSTRUCTION = (
    "Explain your reasoning step by step using the relevant visual evidence and any text in the image."
)
DEFAULT_FINAL_ANSWER_INSTRUCTION = "Provide the final answer only."
REPO_ROOT = Path(__file__).resolve().parents[2]


def ensure_internvl2_5_template() -> None:
    try:
        get_conv_template("internvl2_5")
        return
    except KeyError:
        pass

    register_conv_template(
        Conversation(
            name="internvl2_5",
            system_template="<|im_start|>system\n{system_message}",
            system_message="你是书生·万象，英文名是InternVL，是由上海人工智能实验室、清华大学及多家合作单位联合开发的多模态大语言模型。",
            roles=("<|im_start|>user\n", "<|im_start|>assistant\n"),
            sep_style=SeparatorStyle.MPT,
            sep="<|im_end|>\n",
        )
    )


def configure_model(checkpoint_path: str) -> InternVLChatConfig:
    config = InternVLChatConfig.from_json_file(os.path.join(checkpoint_path, "config.json"))
    match = re.search(r"InternVL2-(\d+B)", checkpoint_path)
    model_size = match.group(1) if match else checkpoint_path.split("-")[-1]
    if model_size in ["1B", "40B"]:
        config.llm_config._attn_implementation = "eager"
    else:
        config.llm_config.attn_implementation = "eager"
    config.vision_config.use_flash_attn = True
    return config


def split_model(num_layers: int, gpus_per_model: int) -> dict:
    if gpus_per_model < 1:
        raise ValueError("gpus_per_model must be >= 1")

    device_map = {}
    if gpus_per_model == 1:
        for layer_idx in range(num_layers):
            device_map[f"language_model.model.layers.{layer_idx}"] = 0
    else:
        # Keep the vision tower and embeddings on GPU 0 and spread decoder layers.
        num_layers_per_gpu = math.ceil(num_layers / (gpus_per_model - 0.5))
        num_layers_per_gpu = [num_layers_per_gpu] * gpus_per_model
        num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)

        layer_cnt = 0
        for gpu_idx, layer_count in enumerate(num_layers_per_gpu):
            for _ in range(layer_count):
                if layer_cnt >= num_layers:
                    break
                device_map[f"language_model.model.layers.{layer_cnt}"] = gpu_idx
                layer_cnt += 1

        if layer_cnt < num_layers:
            for layer_idx in range(layer_cnt, num_layers):
                device_map[f"language_model.model.layers.{layer_idx}"] = gpus_per_model - 1

    device_map["vision_model"] = 0
    device_map["mlp1"] = 0
    device_map["language_model.model.tok_embeddings"] = 0
    device_map["language_model.model.rotary_emb"] = 0
    device_map["language_model.model.embed_tokens"] = 0
    device_map["language_model.output"] = 0
    device_map["language_model.model.norm"] = 0
    device_map["language_model.lm_head"] = 0
    if num_layers > 1 and gpus_per_model > 1:
        device_map[f"language_model.model.layers.{num_layers - 1}"] = 1
    return device_map


def load_model(checkpoint_path: str, config: InternVLChatConfig, args) -> InternVLChatModel:
    ensure_internvl2_5_template()
    kwargs = {"device_map": "auto"} if args.auto else {}
    if args.gpus_per_model > 1 and not args.auto:
        if args.gpus_per_model > torch.cuda.device_count():
            raise ValueError(
                f"gpus_per_model={args.gpus_per_model} exceeds visible CUDA devices={torch.cuda.device_count()}"
            )
        kwargs["device_map"] = split_model(config.llm_config.num_hidden_layers, args.gpus_per_model)

    model = InternVLChatModel.from_pretrained(
        checkpoint_path,
        config=config,
        low_cpu_mem_usage=True,
        torch_dtype=torch.bfloat16,
        load_in_8bit=args.load_in_8bit,
        load_in_4bit=args.load_in_4bit,
        **kwargs,
    ).eval()

    if args.gpus_per_model == 1 and not args.auto and not args.load_in_8bit and not args.load_in_4bit:
        model = model.cuda()
    return model


def dataset_prompt(dataset_name: str) -> str:
    if "vizwiz" in dataset_name:
        return VIZWIZ_PROMPT + BASE_PROMPT
    if "ai2d" in dataset_name:
        return AI2D_PROMPT
    if "infographicsvqa" in dataset_name:
        return INFOGRAPHICSVQA_PROMPT
    return BASE_PROMPT


def resolve_dataset_path(data_root: str, path: str) -> str:
    if os.path.isabs(path):
        return path
    return os.path.join(data_root, path)


def resolve_image_path(image_path: str, data_root: str, jsonl_path: str = "") -> str:
    candidates = []
    if os.path.isabs(image_path):
        candidates.append(image_path)

    jsonl_dir = os.path.dirname(jsonl_path) if jsonl_path else ""
    candidates.append(os.path.join(data_root, image_path))
    if image_path.startswith("data/"):
        candidates.append(os.path.join(data_root, image_path[len("data/"):]))
    if jsonl_dir:
        candidates.append(os.path.join(jsonl_dir, image_path))
        candidates.append(os.path.join(jsonl_dir, os.path.basename(image_path)))

    for candidate in candidates:
        if candidate and os.path.exists(candidate):
            return candidate
    raise FileNotFoundError(f"Could not resolve image path: {image_path}")


def load_textvqa_sample(jsonl_path: str, sample_index: int) -> Tuple[str, str, Optional[int], Optional[str]]:
    with open(jsonl_path) as f:
        for idx, line in enumerate(f):
            if idx == sample_index:
                item = json.loads(line)
                return item["image"], item["question"], item.get("question_id"), item.get("answer")
    raise IndexError(f"sample_index {sample_index} is out of range for {jsonl_path}")


def build_pixel_values(
    image_path: str,
    image_size: int,
    dynamic: bool,
    use_thumbnail: bool,
    max_num: int,
) -> torch.Tensor:
    transform = build_transform(is_train=False, input_size=image_size)
    image = Image.open(image_path).convert("RGB")
    if dynamic:
        images = dynamic_preprocess(
            image,
            image_size=image_size,
            use_thumbnail=use_thumbnail,
            max_num=max_num,
        )
    else:
        images = [image]
    return torch.stack([transform(img) for img in images])


def build_query(model: InternVLChatModel, tokenizer, question: str, num_patches: int):
    img_context_token = "<IMG_CONTEXT>"
    img_start_token = "<img>"
    img_end_token = "</img>"

    if "<image>" not in question:
        question = "<image>\n" + question

    model.img_context_token_id = tokenizer.convert_tokens_to_ids(img_context_token)

    template = get_conv_template(model.template)
    template.system_message = model.system_message
    template.append_message(template.roles[0], question)
    template.append_message(template.roles[1], None)
    query = template.get_prompt()

    image_tokens = img_start_token + img_context_token * model.num_image_token * num_patches + img_end_token
    query = query.replace("<image>", image_tokens, 1)
    return query, template


def model_input_device(model: InternVLChatModel) -> torch.device:
    return next(model.vision_model.parameters()).device


@torch.inference_mode()
def generate_answer(
    model: InternVLChatModel,
    tokenizer,
    pixel_values: torch.Tensor,
    question: str,
    generation_config: dict,
) -> str:
    query, template = build_query(model, tokenizer, question, pixel_values.shape[0])
    model_inputs = tokenizer(query, return_tensors="pt")

    device = model_input_device(model)
    input_ids = model_inputs["input_ids"].to(device)
    attention_mask = model_inputs["attention_mask"].to(device)
    eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)

    output_ids = model.generate(
        pixel_values=pixel_values.to(device=device, dtype=torch.bfloat16),
        input_ids=input_ids,
        attention_mask=attention_mask,
        large_model=True,
        eos_token_id=eos_token_id,
        **generation_config,
    )
    response = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0]
    return response.split(template.sep)[0].strip()


def build_eval_entries(result_items, annotation_file: str):
    evaluator = TextVQAAccuracyEvaluator()
    with open(annotation_file) as f:
        annotations = json.load(f)["annotations"]
    question_id_to_answers = {
        item["question_id"]: [answer["answer"] for answer in item["answers"]]
        for item in annotations
    }
    eval_entries = [
        {
            "question_id": item["question_id"],
            "answer": item["answer"],
            "pred_answer": item["answer"],
            "gt_answers": question_id_to_answers[item["question_id"]],
        }
        for item in result_items
    ]
    return evaluator, eval_entries


def make_generation_config(num_beams: int, max_new_tokens: int, temperature: float) -> dict:
    generation_config = {
        "num_beams": num_beams,
        "max_new_tokens": max_new_tokens,
        "min_new_tokens": 1,
        "do_sample": temperature > 0,
    }
    if temperature > 0:
        generation_config["temperature"] = temperature
    return generation_config


def append_instruction(question: str, instruction: str) -> str:
    instruction = instruction.strip()
    if not instruction:
        return question
    return f"{question.rstrip()}\n{instruction}"


def render_custom_prompt(question: str, prompt_template: str) -> str:
    prompt_template = prompt_template.strip()
    if not prompt_template:
        raise ValueError("custom_prompt_template must be non-empty when reasoning_mode='custom_prompt'.")
    if "{question}" in prompt_template:
        return prompt_template.replace("{question}", question)
    if "Question:" in prompt_template or "Question：" in prompt_template:
        return f"{prompt_template.rstrip()} {question}"
    return f"{prompt_template.rstrip()}\nQuestion: {question}"


def extract_final_answer(response: str, final_answer_prefix: str) -> str:
    final_answer_prefix = final_answer_prefix.strip()
    if not final_answer_prefix:
        return response.strip()

    pattern = re.compile(rf"(?im)^{re.escape(final_answer_prefix)}\s*(.*)$")
    match = pattern.search(response)
    if not match:
        return response.strip()

    inline_answer = match.group(1).strip()
    if inline_answer:
        return inline_answer

    trailing_lines = response[match.end():].splitlines()
    for line in trailing_lines:
        stripped = line.strip()
        if stripped:
            return stripped
    return ""


def make_reasoning_generation_config(base_generation_config: dict, args) -> dict:
    generation_config = dict(base_generation_config)
    generation_config["max_new_tokens"] = args.reasoning_max_new_tokens
    temperature = args.reasoning_temperature
    generation_config["do_sample"] = temperature > 0
    if temperature > 0:
        generation_config["temperature"] = temperature
    else:
        generation_config.pop("temperature", None)
    return generation_config


def generate_answer_with_reasoning(
    model: InternVLChatModel,
    tokenizer,
    pixel_values: torch.Tensor,
    question: str,
    generation_config: dict,
    reasoning_mode: str,
    reasoning_generation_config: Optional[dict] = None,
    final_answer_instruction: str = "",
) -> Tuple[str, Optional[str]]:
    if reasoning_mode == "none":
        return generate_answer(model, tokenizer, pixel_values, question, generation_config), None

    if reasoning_mode == "prompt":
        prompted_question = append_instruction(question, HIDDEN_REASONING_INSTRUCTION)
        return generate_answer(model, tokenizer, pixel_values, prompted_question, generation_config), None

    if reasoning_generation_config is None:
        raise ValueError("reasoning_generation_config is required when reasoning_mode='two_pass'.")

    reasoning_question = append_instruction(question, EXPLICIT_REASONING_INSTRUCTION)
    reasoning = generate_answer(model, tokenizer, pixel_values, reasoning_question, reasoning_generation_config)
    final_instruction = final_answer_instruction or DEFAULT_FINAL_ANSWER_INSTRUCTION
    final_question = append_instruction(
        question,
        f"Reasoning:\n{reasoning}\n{final_instruction}",
    )
    answer = generate_answer(model, tokenizer, pixel_values, final_question, generation_config)
    return answer, reasoning


def run_single(args):
    tokenizer = AutoTokenizer.from_pretrained(
        args.checkpoint,
        trust_remote_code=True,
        use_fast=False,
    )
    config = configure_model(args.checkpoint)
    model = load_model(args.checkpoint, config, args)

    if args.textvqa_jsonl:
        image_path, prompt, question_id, answer = load_textvqa_sample(args.textvqa_jsonl, args.sample_index)
        image_path = resolve_image_path(image_path, args.data_root, args.textvqa_jsonl)
    else:
        image_path = args.image_path
        prompt = args.prompt
        question_id = None
        answer = None

    if not image_path or not prompt:
        raise ValueError("Provide either --image-path and --prompt, or --textvqa-jsonl.")
    if not os.path.exists(image_path):
        raise FileNotFoundError(f"image not found: {image_path}")

    image_size = config.force_image_size or config.vision_config.image_size
    pixel_values = build_pixel_values(
        image_path=image_path,
        image_size=image_size,
        dynamic=args.dynamic,
        use_thumbnail=config.use_thumbnail,
        max_num=args.max_num,
    )

    generation_config = make_generation_config(
        num_beams=args.num_beams,
        max_new_tokens=args.max_new_tokens,
        temperature=args.temperature,
    )
    reasoning_generation_config = None
    if args.reasoning_mode == "two_pass":
        reasoning_generation_config = make_reasoning_generation_config(generation_config, args)
    raw_prediction = None
    if args.reasoning_mode == "custom_prompt":
        raw_prediction = generate_answer(
            model,
            tokenizer,
            pixel_values,
            render_custom_prompt(prompt, args.custom_prompt_template),
            generation_config,
        )
        prediction = (
            extract_final_answer(raw_prediction, args.final_answer_prefix)
            if args.extract_final_answer
            else raw_prediction
        )
        reasoning = None
    else:
        prediction, reasoning = generate_answer_with_reasoning(
            model=model,
            tokenizer=tokenizer,
            pixel_values=pixel_values,
            question=prompt,
            generation_config=generation_config,
            reasoning_mode=args.reasoning_mode,
            reasoning_generation_config=reasoning_generation_config,
            final_answer_instruction=args.answer_format_prompt,
        )

    print(f"checkpoint: {args.checkpoint}")
    print(f"image_path: {image_path}")
    if question_id is not None:
        print(f"question_id: {question_id}")
    if answer is not None:
        print(f"reference_answer: {answer}")
    print(f"prompt: {prompt}")
    if reasoning is not None:
        print(f"reasoning: {reasoning}")
    if raw_prediction is not None:
        print(f"raw_prediction: {raw_prediction}")
    print(f"prediction: {prediction}")


def run_textvqa_eval(args):
    if args.dataset not in ds_collections:
        raise KeyError(f"unknown dataset: {args.dataset}")

    ds_cfg = ds_collections[args.dataset]
    test_file = args.test_file or resolve_dataset_path(args.data_root, ds_cfg["test"])
    train_file = args.train_file or resolve_dataset_path(args.data_root, ds_cfg["train"])
    annotation_file = args.annotation_file or resolve_dataset_path(args.data_root, ds_cfg["annotation"])

    tokenizer = AutoTokenizer.from_pretrained(
        args.checkpoint,
        trust_remote_code=True,
        use_fast=False,
    )
    config = configure_model(args.checkpoint)
    model = load_model(args.checkpoint, config, args)

    image_size = config.force_image_size or config.vision_config.image_size
    prompt = args.prompt or dataset_prompt(args.dataset)
    dataset = VQADataset(
        train=train_file,
        test=test_file,
        prompt=prompt,
        few_shot=0,
        input_size=image_size,
        dynamic_image_size=args.dynamic,
        use_thumbnail=config.use_thumbnail,
        max_num=args.max_num,
    )

    num_items = len(dataset) if args.limit is None else min(len(dataset), args.limit)
    result_items = []
    generation_config = make_generation_config(
        num_beams=args.num_beams,
        max_new_tokens=args.max_new_tokens or ds_cfg["max_new_tokens"],
        temperature=args.temperature,
    )
    reasoning_generation_config = None
    if args.reasoning_mode == "two_pass":
        reasoning_generation_config = make_reasoning_generation_config(generation_config, args)

    for idx in range(num_items):
        sample = dataset[idx]
        raw_prediction = None
        if args.reasoning_mode == "custom_prompt":
            raw_prediction = generate_answer(
                model,
                tokenizer,
                sample["pixel_values"],
                render_custom_prompt(sample["question"], args.custom_prompt_template),
                generation_config,
            )
            prediction = (
                extract_final_answer(raw_prediction, args.final_answer_prefix)
                if args.extract_final_answer
                else raw_prediction
            )
            reasoning = None
        else:
            prediction, reasoning = generate_answer_with_reasoning(
                model=model,
                tokenizer=tokenizer,
                pixel_values=sample["pixel_values"],
                question=sample["question"],
                generation_config=generation_config,
                reasoning_mode=args.reasoning_mode,
                reasoning_generation_config=reasoning_generation_config,
            )
        result_item = {
            "question": sample["question"],
            "question_id": sample["question_id"],
            "answer": prediction,
            "annotation": sample["annotation"],
        }
        if raw_prediction is not None:
            result_item["raw_answer"] = raw_prediction
        if args.save_reasoning and reasoning is not None:
            result_item["reasoning"] = reasoning
        result_items.append(result_item)
        if (idx + 1) % args.log_every == 0 or idx + 1 == num_items:
            print(f"[{idx + 1}/{num_items}] question_id={sample['question_id']} prediction={prediction}")
            sys.stdout.flush()

    os.makedirs(args.out_dir, exist_ok=True)
    run_name = args.run_name or f"{args.dataset}_{os.path.basename(args.checkpoint)}"
    output_file = os.path.join(args.out_dir, f"{run_name}.json")
    with open(output_file, "w") as f:
        json.dump(result_items, f, ensure_ascii=False, indent=2)

    evaluator, eval_entries = build_eval_entries(result_items, annotation_file)
    accuracy = evaluator.eval_pred_list(eval_entries)
    print(f"dataset: {args.dataset}")
    print(f"checkpoint: {args.checkpoint}")
    print(f"count: {num_items}")
    print(f"accuracy: {accuracy:.6f}")
    print(f"results_file: {output_file}")


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--checkpoint", type=str, required=True)
    parser.add_argument("--mode", type=str, choices=["single", "textvqa_eval"], default="single")
    parser.add_argument("--image-path", type=str, default="")
    parser.add_argument("--prompt", type=str, default="")
    parser.add_argument("--textvqa-jsonl", type=str, default="")
    parser.add_argument("--sample-index", type=int, default=0)
    parser.add_argument("--dataset", type=str, default="textvqa_val")
    parser.add_argument("--data-root", type=str, default=str(REPO_ROOT))
    parser.add_argument("--test-file", type=str, default="")
    parser.add_argument("--train-file", type=str, default="")
    parser.add_argument("--annotation-file", type=str, default="")
    parser.add_argument("--out-dir", type=str, default=str(REPO_ROOT / "outputs" / "native_single"))
    parser.add_argument("--run-name", type=str, default="")
    parser.add_argument("--limit", type=int, default=None)
    parser.add_argument("--max-new-tokens", type=int, default=0)
    parser.add_argument("--num-beams", type=int, default=1)
    parser.add_argument("--temperature", type=float, default=0.0)
    parser.add_argument("--reasoning-mode", type=str, choices=["none", "prompt", "two_pass", "custom_prompt"], default="none")
    parser.add_argument("--reasoning-max-new-tokens", type=int, default=64)
    parser.add_argument("--reasoning-temperature", type=float, default=0.0)
    parser.add_argument("--save-reasoning", action="store_true")
    parser.add_argument("--answer-format-prompt", type=str, default="")
    parser.add_argument("--custom-prompt-template", type=str, default="")
    parser.add_argument("--extract-final-answer", action="store_true")
    parser.add_argument("--final-answer-prefix", type=str, default="Final answer:")
    parser.add_argument("--dynamic", action="store_true")
    parser.add_argument("--max-num", type=int, default=6)
    parser.add_argument("--log-every", type=int, default=20)
    parser.add_argument("--seed", type=int, default=0)
    parser.add_argument("--gpus-per-model", type=int, default=1)
    parser.add_argument("--auto", action="store_true")
    parser.add_argument("--load-in-8bit", action="store_true")
    parser.add_argument("--load-in-4bit", action="store_true")
    args = parser.parse_args()

    if not torch.cuda.is_available():
        raise RuntimeError("CUDA is required for native InternVL inference.")

    random.seed(args.seed)
    torch.manual_seed(args.seed)

    if args.mode == "single":
        if args.max_new_tokens == 0:
            args.max_new_tokens = 32
        run_single(args)
        return

    if args.max_new_tokens == 0:
        args.max_new_tokens = None
    run_textvqa_eval(args)


if __name__ == "__main__":
    main()