llava.py

import logging
import warnings
from typing import List, Optional, Tuple, Union

import torch
from accelerate import Accelerator, DistributedType
from accelerate.state import AcceleratorState
from tqdm import tqdm

from lm_eval.api.instance import Instance
from lm_eval.api.model import LM
from lm_eval.api.registry import register_model
from lm_eval.models.utils import Collator


warnings.filterwarnings("ignore")

eval_logger = logging.getLogger("lm-eval")

try:
    from llava.constants import (
        DEFAULT_IMAGE_TOKEN,
        IMAGE_TOKEN_INDEX,
    )
    from llava.conversation import conv_templates
    from llava.mm_utils import (
        get_model_name_from_path,
        process_images,
        tokenizer_image_token,
    )
    from llava.model.builder import load_pretrained_model
except ImportError:
    eval_logger.error("LLaVA is not installed. Please install LLaVA to use this model.")


@register_model("llava")
class Llava(LM):
    """
    Llava Model
    """

    def __init__(
        self,
        pretrained: str = "liuhaotian/llava-v1.5-7b",
        truncation: Optional[bool] = True,
        device: Optional[str] = "cuda",
        dtype: Optional[Union[str, torch.dtype]] = "auto",
        batch_size: Optional[Union[int, str]] = 1,
        trust_remote_code: Optional[bool] = False,
        revision=None,
        use_flash_attention_2=False,
        conv_template="vicuna_v1",
        use_cache=True,
        truncate_context=False,  # whether to truncate the context in generation, set it False for LLaVA-1.6
        **kwargs,
    ) -> None:
        super().__init__()
        # Do not use kwargs for now
        assert kwargs == {}, f"Unexpected kwargs: {kwargs}"

        accelerator = Accelerator()
        if accelerator.num_processes > 1:
            self._device = torch.device(f"cuda:{accelerator.local_process_index}")
        else:
            self._device = device
        (
            self._tokenizer,
            self._model,
            self._image_processor,
            self._max_length,
        ) = load_pretrained_model(
            pretrained,
            None,
            get_model_name_from_path(pretrained),
            device_map=self._device,
        )
        self._config = self._model.config
        self.model.eval()
        self.model.tie_weights()
        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
        else:
            self.model.to(self._device)
            self._rank = 0
            self._word_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):
        return self.tokenizer.decode(tokens)

    def loglikelihood(self, requests: List[Instance]) -> List[Tuple[float, bool]]:
        # TODO
        res = []
        pbar = tqdm(
            total=len(requests), disable=(self.rank != 0), desc="Model Responding"
        )

        print(requests[0])
        import sys; sys.exit()
        for contexts, doc_to_target, doc_to_visual, doc, task in [
            reg.args for reg in requests
        ]:
            # encode, pad, and truncate contexts for this batch
            if isinstance(doc_to_target, str):
                continuation = doc_to_target
            else:
                continuation = doc_to_target(doc)
            visuals = [doc_to_visual(doc)]
            visuals = self.flatten(visuals)
            if visuals:
                image = process_images(visuals, self._image_processor, self._config)
                if isinstance(image, list):
                    image = [
                        _image.to(dtype=torch.float16, device=self.device)
                        for _image in image
                    ]
                else:
                    image = image.to(dtype=torch.float16, device=self.device)
            else:
                image = None

            prompts_input = contexts[0]

            if (
                image is not None
                and len(image) != 0
                and DEFAULT_IMAGE_TOKEN not in prompts_input
            ):
                """
                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.
                """
                image_tokens = [DEFAULT_IMAGE_TOKEN] * len(visuals)
                image_tokens = " ".join(image_tokens)
                prompts_input = image_tokens + "\n" + contexts[0]

            conv = conv_templates[self.conv_template].copy()
            conv.append_message(conv.roles[0], prompts_input)
            conv.append_message(conv.roles[1], None)
            prompt = conv.get_prompt()
            contxt_id = (
                tokenizer_image_token(
                    prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
                )
                .unsqueeze(0)
                .to(self.device)
            )
            # Add the answer of the second role
            conv.messages[1][1] = continuation

            prompt = conv.get_prompt()
            input_ids = (
                tokenizer_image_token(
                    prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt"
                )
                .unsqueeze(0)
                .to(self.device)
            )
            labels = input_ids.clone()
            # Context part no need to calculate for loss
            labels[0, : contxt_id.shape[1]] = -100
            with torch.inference_mode():
                outputs = self.model(
                    input_ids=input_ids, labels=labels, images=image, use_cache=True
                )
            loss = outputs["loss"]
            # loss = torch.exp(loss)
            logits = outputs["logits"]
            greedy_tokens = logits.argmax(dim=-1)
            cont_toks = input_ids[:, contxt_id.shape[1] :]  # [1, seq]
            greedy_tokens = greedy_tokens[
                :, contxt_id.shape[1] : input_ids.shape[1]
            ]  # [1, seq]
            max_equal = (greedy_tokens == cont_toks).all()
            res.append((float(loss.item()), bool(max_equal)))
            pbar.update(1)
        pbar.close()
        return res

    def loglikelihood_rolling(
        self, requests: List[Instance]
    ) -> List[Tuple[float, bool]]:
        raise NotImplementedError()

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

    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.
        re_ords = Collator(
            [reg.arguments for reg in requests],
            sort_fn=_collate,
            group_by="gen_kwargs",
            group_fn=lambda x: x[1],
        )
        chunks = re_ords.get_batched(n=self.batch_size)
        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:
            contexts, all_gen_kwargs, visuals = zip(*chunk)
            # task = task[0]

            # visuals = [doc_to_visual[0](doc[0])]
            visuals = self.flatten(visuals)
            # 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]

            # Set default values for until and max_new_tokens
            until = [self.tok_decode(self.eot_token_id)]

            # Update values from gen_kwargs if present
            if "until" in gen_kwargs:
                until = gen_kwargs.pop("until")
                if isinstance(until, str):
                    until = [until]
                elif not isinstance(until, list):
                    raise ValueError(
                        f"Expected `gen_kwargs['until']` to be of type Union[str,list] but got {type(until)}"
                    )

            if (
                "image_aspect_ratio" in gen_kwargs.keys()
                and "image_aspect_ratio" not in self._config.__dict__
            ):
                # here we should pop it out of gen_kwargs so that it doesn't get passed to the model for next step of generation
                self._config.image_aspect_ratio = gen_kwargs.pop("image_aspect_ratio")
                eval_logger.info(
                    f"Setting image aspect ratio: {self._config.image_aspect_ratio}"
                )
            # encode, pad, and truncate contexts for this batch
            if visuals:
                image_tensor = process_images(
                    visuals, self._image_processor, self._config
                )
                if isinstance(image_tensor, list):
                    image_tensor = [
                        _image.to(dtype=torch.float16, device=self.device)
                        for _image in image_tensor
                    ]
                else:
                    image_tensor = image_tensor.to(
                        dtype=torch.float16, device=self.device
                    )
            else:
                image_tensor = None

            # prompts_input = contexts[0]

            question_input = []

            for visual, context in zip(visuals, contexts):
                if (
                    image_tensor is not None
                    and len(image_tensor) != 0
                    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.
                    """
                    image_tokens = (
                        [DEFAULT_IMAGE_TOKEN] * len(visual)
                        if isinstance(visual, list)
                        else [DEFAULT_IMAGE_TOKEN]
                    )
                    image_tokens = " ".join(image_tokens)
                    question = image_tokens + "\n" + context
                else:
                    question = context

                conv = conv_templates[self.conv_template].copy()
                conv.append_message(conv.roles[0], question)
                conv.append_message(conv.roles[1], None)
                prompt_question = conv.get_prompt()
                question_input.append(prompt_question)

            # The above for loop has bugs. When there is no visuals, e.g. pure text,
            # there will be no for loop execute resulting in an empty question_input (because no visuals)
            # Scenario 1 won't even be execute
            if len(visuals) == 0:
                for context in contexts:
                    question = context
                    conv = conv_templates[self.conv_template].copy()
                    conv.append_message(conv.roles[0], question)
                    conv.append_message(conv.roles[1], None)
                    prompt_question = conv.get_prompt()
                    question_input.append(prompt_question)

            # input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(self.device)
            # preconfigure gen_kwargs with defaults
            gen_kwargs["image_sizes"] = [
                visuals[idx].size for idx in range(len(visuals))
            ]
            if "max_gen_toks" not in gen_kwargs:
                gen_kwargs["max_gen_toks"] = 1024
            if "temperature" not in gen_kwargs:
                gen_kwargs["temperature"] = 0
            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)
            # These steps are not in LLaVA's original code, but are necessary for generation to work
            # TODO: pay attention to this major generation step...
            try:
                cont = self.model.generate(
                    input_ids,
                    attention_mask=attention_masks,
                    pad_token_id=pad_token_ids,
                    images=image_tensor,
                    image_sizes=gen_kwargs["image_sizes"],
                    do_sample=gen_kwargs["do_sample"],
                    temperature=gen_kwargs["temperature"],
                    top_p=gen_kwargs["top_p"],
                    num_beams=gen_kwargs["num_beams"],
                    max_new_tokens=gen_kwargs["max_gen_toks"],
                    use_cache=self.use_cache,
                )
                text_outputs = self.tokenizer.batch_decode(
                    cont, skip_special_tokens=True
                )
            except Exception as e:
                eval_logger.error(f"Error {e} in generating")
                cont = ""
                text_outputs = [""]

            # cont_toks_list = cont.tolist()
            # for cont_toks, context in zip(cont_toks_list, contexts):
            # discard context + left-padding toks if using causal decoder-only LMM
            # if self.truncate_context:
            #     cont_toks = cont_toks[input_ids.shape[1] :]
            # use secondary stop seqs to cut off should-have-been-stopped content post-hoc
            # if self.truncate_context:
            #     for term in until:
            #         if len(term) > 0:
            #             # ignore '' separator,
            #             # for seq2seq case where self.tok_decode(self.eot_token_id) = ''
            #             text_outputs = text_outputs.split(term)[0]
            res.extend(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