src/lmms-eval/lmms_eval/models/simple/llava_onevision_moviechat.py

import copy
import json
import logging
import math
import os
import re
import warnings
from datetime import timedelta
from typing import List, Optional, Tuple, Union

import av
import numpy as np
import PIL
import torch
import transformers
from accelerate import Accelerator, DistributedType, InitProcessGroupKwargs
from accelerate.state import AcceleratorState
from decord import VideoReader, cpu
from moviepy.video.io.VideoFileClip import VideoFileClip
from packaging import version
from PIL import Image
from tqdm import tqdm
from transformers import AutoConfig

from lmms_eval import utils
from lmms_eval.api.instance import Instance
from lmms_eval.api.model import lmms
from lmms_eval.api.registry import register_model

# Suppress warnings
warnings.filterwarnings("ignore")

# Configure logging
eval_logger = logging.getLogger("lmms-eval")

# Enable TF32 for CUDA
torch.backends.cuda.matmul.allow_tf32 = True

# Import LLaVA modules
try:
    from llava.constants import (
        DEFAULT_IM_END_TOKEN,
        DEFAULT_IM_START_TOKEN,
        DEFAULT_IMAGE_TOKEN,
        IGNORE_INDEX,
        IMAGE_TOKEN_INDEX,
    )
    from llava.conversation import SeparatorStyle, conv_templates
    from llava.mm_utils import (
        KeywordsStoppingCriteria,
        get_model_name_from_path,
        process_images,
        tokenizer_image_token,
    )
    from llava.model.builder import load_pretrained_model
except ImportError as e:
    eval_logger.debug(f"LLaVA_NeXT is not installed. Please install llava from `https://github.com/rese1f/MovieChat.git` to use this model.\nError: {e}")


# Determine best attention implementation
if version.parse(torch.__version__) >= version.parse("2.1.2"):
    best_fit_attn_implementation = "sdpa"
else:
    best_fit_attn_implementation = "eager"


# llava_onevision_moviechat uses the same memory consolidation mechanism with the original MovieChat, but changes the base model from VideoLLamA to LLava-OneVision
@register_model("llava_onevision_moviechat")
class Llava_OneVision_MovieChat(lmms):
    """
    Llava Model
    """

    def __init__(
        self,
        pretrained: str = "lmms-lab/llava-onevision-qwen2-7b-ov",
        truncation: Optional[bool] = True,
        device: Optional[str] = "cuda:0",
        batch_size: Optional[Union[int, str]] = 1,
        model_name: str = "llava_qwen",
        attn_implementation: Optional[str] = best_fit_attn_implementation,
        device_map: Optional[str] = "cuda:0",
        conv_template: Optional[str] = "qwen_1_5",
        use_cache: Optional[bool] = True,
        truncate_context: Optional[bool] = False,  # whether to truncate the context in generation, set it False for LLaVA-1.6
        customized_config: Optional[str] = None,  # ends in json
        short_memory_length: Optional[int] = 18,
        long_memory_length: Optional[int] = 64,
        sliding_window_length: Optional[int] = 8,
        merge_frame_length: Optional[int] = 2,
        tmp_folder: Optional[str] = "tmp/",
        mm_spatial_pool_stride: Optional[int] = 2,
        mm_spatial_pool_mode: Optional[str] = "bilinear",
        token_strategy: Optional[str] = "single",  # could be "single" or "multiple", "multiple" denotes adding multiple <image> tokens for each frame
        video_decode_backend: str = "decord",
        **kwargs,
    ) -> None:
        super().__init__()
        # Do not use kwargs for now
        assert kwargs == {}, f"Unexpected kwargs: {kwargs}"

        accelerator_kwargs = InitProcessGroupKwargs(timeout=timedelta(weeks=52))
        accelerator = Accelerator(kwargs_handlers=[accelerator_kwargs])
        if accelerator.num_processes > 1:
            self._device = torch.device(f"cuda:{accelerator.local_process_index}")
            self.device_map = f"cuda:{accelerator.local_process_index}"
        elif accelerator.num_processes == 1 and device_map == "auto":
            self._device = torch.device(device)
            self.device_map = device_map
        else:
            self._device = torch.device(f"cuda:{accelerator.local_process_index}")
            self.device_map = f"cuda:{accelerator.local_process_index}"

        llava_model_args = {
            "multimodal": True,
        }
        if customized_config is not None:
            llava_model_args["customized_config"] = customized_config
        if attn_implementation is not None:
            llava_model_args["attn_implementation"] = attn_implementation
        if "use_flash_attention_2" in kwargs:
            llava_model_args["use_flash_attention_2"] = kwargs["use_flash_attention_2"]
        model_name = model_name if model_name is not None else get_model_name_from_path(pretrained)

        self.pretrained = pretrained
        self.token_strategy = token_strategy
        self.mm_spatial_pool_stride = mm_spatial_pool_stride
        self.mm_spatial_pool_mode = mm_spatial_pool_mode
        self.video_decode_backend = video_decode_backend

        self.short_memory_length = short_memory_length
        self.long_memory_length = long_memory_length
        self.merge_frame_length = merge_frame_length
        self.sliding_window_length = sliding_window_length
        self.num_clips = (self.long_memory_length // self.merge_frame_length) * ((self.short_memory_length - self.merge_frame_length) // self.sliding_window_length)
        self.tmp_folder = tmp_folder

        overwrite_config = {}
        overwrite_config["mm_spatial_pool_stride"] = self.mm_spatial_pool_stride
        overwrite_config["mm_spatial_pool_mode"] = self.mm_spatial_pool_mode
        cfg_pretrained = AutoConfig.from_pretrained(self.pretrained)

        if cfg_pretrained.architectures[0] == "LlavaLlamaForCausalLM":  # Ugly code, only used in  vicuna that needs ROPE
            if "224" in cfg_pretrained.mm_vision_tower:
                least_token_number = self.max_frames_num * (16 // self.mm_spatial_pool_stride) ** 2 + 1000
            else:
                least_token_number = self.max_frames_num * (24 // self.mm_spatial_pool_stride) ** 2 + 1000

            scaling_factor = math.ceil(least_token_number / 4096)
            if scaling_factor >= 2:
                overwrite_config["rope_scaling"] = {"factor": float(scaling_factor), "type": "linear"}
                overwrite_config["max_sequence_length"] = 4096 * scaling_factor
                overwrite_config["tokenizer_model_max_length"] = 4096 * scaling_factor

        llava_model_args["overwrite_config"] = overwrite_config
        from LLaVA_NeXT.llava.model.builder import load_pretrained_model

        try:
            # Try to load the model with the multimodal argument
            self._tokenizer, self._model, self._image_processor, self._max_length = load_pretrained_model(pretrained, None, model_name, device_map=self.device_map, **llava_model_args)
        except TypeError:
            # for older versions of LLaVA that don't have multimodal argument
            llava_model_args.pop("multimodal", None)
            self._tokenizer, self._model, self._image_processor, self._max_length = load_pretrained_model(pretrained, None, model_name, device_map=self.device_map, **llava_model_args)

        self._config = self._model.config
        self.model.eval()
        self.truncation = truncation
        self.batch_size_per_gpu = int(batch_size)
        self.conv_template = conv_template
        self.use_cache = use_cache
        self.truncate_context = truncate_context
        assert self.batch_size_per_gpu == 1, "Llava currently does not support batched generation. See https://github.com/haotian-liu/LLaVA/issues/754. HF Llava also has this issue."

        if accelerator.num_processes > 1:
            assert accelerator.distributed_type in [DistributedType.FSDP, DistributedType.MULTI_GPU, DistributedType.DEEPSPEED], "Unsupported distributed type provided. Only DDP and FSDP are supported."
            # If you want to use DistributedType.DEEPSPEED, you have to run accelerate config before using the model
            # Also, you have to select zero stage 0 (equivalent to DDP) in order to make the prepare model works
            # I tried to set different parameters in the kwargs to let default zero 2 stage works, but it didn't work.
            if accelerator.distributed_type == DistributedType.DEEPSPEED:
                kwargs = {
                    "train_micro_batch_size_per_gpu": self.batch_size_per_gpu,
                    "train_batch_size": self.batch_size_per_gpu * accelerator.num_processes,
                }
                AcceleratorState().deepspeed_plugin.deepspeed_config_process(must_match=True, **kwargs)
                eval_logger.info("Detected that you are using DistributedType.DEEPSPEED. Make sure you run `accelerate config` and set zero stage to 0")

            if accelerator.distributed_type == DistributedType.FSDP or accelerator.distributed_type == DistributedType.DEEPSPEED:
                self._model = accelerator.prepare(self.model)
            else:
                self._model = accelerator.prepare_model(self.model, evaluation_mode=True)
            self.accelerator = accelerator
            if self.accelerator.is_local_main_process:
                eval_logger.info(f"Using {accelerator.num_processes} devices with data parallelism")
            self._rank = self.accelerator.local_process_index
            self._world_size = self.accelerator.num_processes

        elif accelerator.num_processes == 1 and device_map == "auto":
            eval_logger.info(f"Using {accelerator.num_processes} devices with tensor parallelism")
            self._rank = 0
            self._world_size = 1

        else:
            eval_logger.info(f"Using single device: {self._device}")
            self.model.to(self._device)
            self._rank = 0
            self._world_size = 1

    @property
    def config(self):
        # return the associated transformers.AutoConfig for the given pretrained model.
        return self._config

    @property
    def tokenizer(self):
        return self._tokenizer

    @property
    def model(self):
        # returns the model, unwrapping it if using Accelerate
        if hasattr(self, "accelerator"):
            return self.accelerator.unwrap_model(self._model)
        else:
            return self._model

    @property
    def eot_token_id(self):
        # we use EOT because end of *text* is more accurate for what we're doing than end of *sentence*
        return self.tokenizer.eos_token_id

    @property
    def max_length(self):
        return self._max_length

    def pad_sequence(self, input_ids, batch_first, padding_value):
        if self.tokenizer.padding_side == "left":
            input_ids = [torch.flip(_input_ids, [0]) for _input_ids in input_ids]
        input_ids = torch.nn.utils.rnn.pad_sequence(input_ids, batch_first=batch_first, padding_value=padding_value)
        if self.tokenizer.padding_side == "left":
            input_ids = torch.flip(input_ids, [1])
        return input_ids

    @property
    def batch_size(self):
        return self.batch_size_per_gpu

    @property
    def device(self):
        return self._device

    @property
    def rank(self):
        return self._rank

    @property
    def world_size(self):
        return self._world_size

    def tok_encode(self, string: str, left_truncate_len=None, add_special_tokens=None) -> List[int]:
        """ """
        add_special_tokens = False if add_special_tokens is None else add_special_tokens
        encoding = self.tokenizer.encode(string, add_special_tokens=add_special_tokens)
        # left-truncate the encoded context to be at most `left_truncate_len` tokens long
        if left_truncate_len:
            encoding = encoding[-left_truncate_len:]
        return encoding

    def tok_decode(self, tokens):
        try:
            return self.tokenizer.decode(tokens)
        except:
            return self.tokenizer.decode([tokens])

    def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
        # TODO
        raise NotImplementedError("MovieChat only supports generation.")

    def flatten(self, input):
        new_list = []
        for i in input:
            for j in i:
                new_list.append(j)
        return new_list

    def load_video(self, video_path, max_frames_num):
        if type(video_path) == str:
            vr = VideoReader(video_path, ctx=cpu(0))
        else:
            vr = VideoReader(video_path[0], ctx=cpu(0))
        total_frame_num = len(vr)
        uniform_sampled_frames = np.linspace(0, total_frame_num - 1, max_frames_num, dtype=int)
        frame_idx = uniform_sampled_frames.tolist()
        spare_frames = vr.get_batch(frame_idx).numpy()
        return spare_frames  # (frames, height, width, channels)

    def extract_keyframes(self, video_path, keyframes):
        container = av.open(video_path)
        video_stream = container.streams.video[0]
        fps = video_stream.average_rate
        time_base = video_stream.time_base
        frames = []

        for keyframe in keyframes:
            keyframe_time = float(keyframe)
            frame_number = int(keyframe_time * fps)
            container.seek(int(keyframe_time / time_base))
            found = False
            for packet in container.demux(video=0):
                for frame in packet.decode():
                    if frame.index >= frame_number:
                        frames.append(frame)
                        found = True
                        break
                if found:
                    break

            if not found:
                container.seek(-1, any_frame=False)
                for packet in container.demux(video=0):
                    for frame in packet.decode():
                        pass
                frames.append(frame)

        video = [x.to_ndarray(format="rgb24") for x in frames]
        video_frames = [Image.fromarray(x) for x in video]
        return video_frames

    def generate_until(self, requests: List[Instance]) -> List[str]:
        res = []

        def _collate(x):
            # the negative sign on len(toks) sorts descending - this has a few advantages:
            # - time estimates will always be over not underestimates, which is more useful for planning
            # - to know the size of a batch when going through the list, you know the first one is always the batch
            #   padded context length. this is useful to simplify the batching logic and more importantly to make
            #   automatic adaptive batches much much easier to implement
            # - any OOMs will happen right away rather than near the end
            toks = self.tok_encode(x[0])
            return -len(toks), x[0]

        # we group requests by their generation_kwargs,
        # so that we don't try to execute e.g. greedy sampling and temp=0.8 sampling
        # in the same batch.
        metadata = requests[0].metadata
        re_ords = utils.Collator([reg.args for reg in requests], _collate, grouping=True)
        chunks = re_ords.get_batched(n=self.batch_size, batch_fn=None)
        num_iters = len(requests) // self.batch_size if len(requests) % self.batch_size == 0 else len(requests) // self.batch_size + 1
        pbar = tqdm(total=num_iters, disable=(self.rank != 0), desc="Model Responding")
        for chunk in chunks:
            batched_contexts, all_gen_kwargs, batched_doc_to_visual, batched_doc_id, batched_task, batched_split = zip(*chunk)
            task = batched_task[0]
            split = batched_split[0]
            batched_visuals = [batched_doc_to_visual[0](self.task_dict[task][split][ids]) for ids in batched_doc_id]  # [B, N]
            assert len(batched_visuals) == 1

            # we assume all gen kwargs in the batch are the same
            # this is safe to assume because the `grouper` object ensures it.
            gen_kwargs = all_gen_kwargs[0]
            if "until" in gen_kwargs:
                gen_kwargs.pop("until")

            question_input = []

            for visual, context in zip(batched_visuals, batched_contexts):
                if len(visual) > 1 or "image_aspect_ratio" not in self._config.__dict__:  # for multi image case, we treat per image aspect ratio as "pad" by default.
                    self._config.image_aspect_ratio = getattr(gen_kwargs, "image_aspect_ratio", "pad")
                    eval_logger.info(f"Setting image aspect ratio: {self._config.image_aspect_ratio}")
                if type(visual[0]) == PIL.Image.Image and "task_type" not in metadata and "sample_frames" not in metadata:  # For image task
                    raise NotImplementedError("MovieChat only supports video inputs.")

                elif "task_type" in metadata and metadata["task_type"] == "video" and "sample_frames" in metadata:
                    raise NotImplementedError("MovieChat only supports video inputs.")

                elif type(visual[0]) == str:  # For video task
                    try:
                        self.short_memory_buffer = []
                        self.long_memory_buffer = []
                        # try:
                        os.makedirs(self.tmp_folder, exist_ok=True)

                        video = VideoFileClip(visual[0])
                        clip_duration = video.duration / self.num_clips

                        cur_frame = 0
                        for i in range(self.num_clips):
                            start_time = i * clip_duration
                            end_time = start_time + clip_duration
                            # uniformly sample self.sliding_window_length frames from the video from start_time to end_time
                            frames = list(video.subclip(start_time, end_time).iter_frames(fps=self.sliding_window_length / clip_duration))[: self.sliding_window_length]
                            frames = [Image.fromarray(frame).convert("RGB") for frame in frames]
                            preprocess_frames = self._image_processor.preprocess(frames, return_tensors="pt")["pixel_values"].half().cuda()
                            encoded_window = self.model.encode_images(preprocess_frames)  # [frames, 729,3584]

                            for frame in encoded_window:
                                if cur_frame < (self.short_memory_length - self.merge_frame_length):
                                    if len(self.short_memory_buffer) == self.short_memory_length:
                                        self.short_memory_buffer.pop(0)
                                    self.short_memory_buffer.append(frame)
                                cur_frame += 1

                            if cur_frame == (self.short_memory_length - self.merge_frame_length):
                                cur_frame = 0

                                # merge short_memory_frames
                                similar_list = []
                                for frame_i in range(len(self.short_memory_buffer) - 1):
                                    scores = self.short_memory_buffer[frame_i] @ self.short_memory_buffer[frame_i + 1].transpose(-1, -2)
                                    frame_silimar = torch.mean(scores)
                                    similar_list.append(frame_silimar)

                                while len(self.short_memory_buffer) > self.merge_frame_length:
                                    max_value = max(similar_list)
                                    max_index = similar_list.index(max_value)
                                    new_frame_feature = (self.short_memory_buffer[max_index].cpu() + self.short_memory_buffer[max_index + 1].cpu()) / 2
                                    self.short_memory_buffer[max_index] = new_frame_feature.cuda()
                                    del self.short_memory_buffer[max_index + 1]
                                    similar_list = []
                                    for frame_i in range(len(self.short_memory_buffer) - 1):
                                        scores = self.short_memory_buffer[frame_i] @ self.short_memory_buffer[frame_i + 1].transpose(-1, -2)
                                        frame_silimar = torch.mean(scores)
                                        similar_list.append(frame_silimar)

                                for frame in self.short_memory_buffer:
                                    self.long_memory_buffer.append(frame)

                        image_features = torch.stack(self.long_memory_buffer)
                    except Exception as e:
                        print(e)
                        eval_logger.error(f"Error {e} in loading video")
                        image_features = None

                    task_type = "video"
                    placeholder_count = len(frames) if self.token_strategy == "multiple" else 1

                if image_features is not None and DEFAULT_IMAGE_TOKEN not in context:
                    """
                    Three senarios:
                    1. No image, and there for, no image token should be added.
                    2. image token is already specified in the context, so we don't need to add it.
                    3. image token is not specified in the context and there is image inputs, so we need to add it. In this case, we add the image token at the beginning of the context and add a new line.
                    4. For video tasks, we could add a <image> token or multiple <image> tokens for each frame in the context. This depends on the training strategy and should balance in test to decide which is better
                    """
                    # if task_type == "image": # indeed in multi-image case, not the video in frames.
                    #     image_tokens = [DEFAULT_IMAGE_TOKEN] * placeholder_count if isinstance(visual, list) else [DEFAULT_IMAGE_TOKEN]
                    # elif task_type == "video":
                    # image_tokens = [DEFAULT_IMAGE_TOKEN] * placeholder_count if self.token_strategy == "multiple" else [DEFAULT_IMAGE_TOKEN]
                    image_tokens = [DEFAULT_IMAGE_TOKEN] * placeholder_count
                    image_tokens = " ".join(image_tokens)
                    question = image_tokens + "\n" + context
                else:
                    question = context

                # This is much safer for llama3, as we now have some object type in it
                if "llama_3" in self.conv_template:
                    conv = copy.deepcopy(conv_templates[self.conv_template])
                else:
                    conv = conv_templates[self.conv_template].copy()

                if utils.is_json(question):  # conversational question input
                    question = json.loads(question)
                    for idx, item in enumerate(question):
                        role = conv.roles[idx % 2]
                        message = item["value"]
                        conv.append_message(role, message)

                    assert len(conv.messages) % 2 == 1
                    conv.append_message(conv.roles[1], None)
                    prompt_question = conv.get_prompt()
                    question_input.append(prompt_question)
                else:  # only simple string for question
                    conv.append_message(conv.roles[0], question)
                    conv.append_message(conv.roles[1], None)
                    prompt_question = conv.get_prompt()
                    question_input.append(prompt_question)

            # preconfigure gen_kwargs with defaults
            if "max_new_tokens" not in gen_kwargs:
                gen_kwargs["max_new_tokens"] = 1024
            if "temperature" not in gen_kwargs:
                gen_kwargs["temperature"] = 0
            if "do_sample" not in gen_kwargs:
                gen_kwargs["do_sample"] = False
            if "top_p" not in gen_kwargs:
                gen_kwargs["top_p"] = None
            if "num_beams" not in gen_kwargs:
                gen_kwargs["num_beams"] = 1

            input_ids_list = [tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt") for prompt in question_input]
            pad_token_ids = self.tokenizer.pad_token_id if self.tokenizer.pad_token_id is not None else self.tokenizer.eos_token_id
            input_ids = self.pad_sequence(input_ids_list, batch_first=True, padding_value=pad_token_ids).to(self.device)
            attention_masks = input_ids.ne(pad_token_ids).to(self.device)

            if task_type == "image":
                gen_kwargs["image_sizes"] = [batched_visuals[0][idx].size for idx in range(len(batched_visuals[0]))]
            elif task_type == "video":
                stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
                keywords = [stop_str]
                stopping_criteria = KeywordsStoppingCriteria(keywords, self.tokenizer, input_ids)
                gen_kwargs["modalities"] = ["video"]
                gen_kwargs["stopping_criteria"] = [stopping_criteria]
                self._config.mm_spatial_pool_stride = self.mm_spatial_pool_stride
                self._config.mm_spatial_pool_mode = self.mm_spatial_pool_mode

            # These steps are not in LLaVA's original code, but are necessary for generation to work
            # TODO: attention to this major generation step...
            if "image_aspect_ratio" in gen_kwargs.keys():
                gen_kwargs.pop("image_aspect_ratio")
            try:
                with torch.inference_mode():
                    gen_kwargs.pop("modalities")
                    cont = self.model.generate_moviechat(input_ids, attention_mask=attention_masks, pad_token_id=pad_token_ids, image_features=image_features, use_cache=self.use_cache, **gen_kwargs)
                text_outputs = self.tokenizer.batch_decode(cont, skip_special_tokens=True)
                text_outputs = [response.strip() for response in text_outputs]
            except Exception as e:
                print(e)
                text_outputs = "Can not infer the answer."

            res.extend(text_outputs)
            print(text_outputs)
            self.cache_hook.add_partial("generate_until", (context, gen_kwargs), text_outputs)
            pbar.update(1)
            # reorder this group of results back to original unsorted form
        res = re_ords.get_original(res)

        pbar.close()
        return res