run_clm.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2020 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.

Here is the full list of checkpoints on the hub that can be fine-tuned by this script:
https://huggingface.co/models?filter=text-generation
"""
# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.

import logging
import math
import os, copy, json
from typing import Sequence, Dict
import sys
from dataclasses import dataclass, field
from itertools import chain
from typing import Optional

import datasets
import evaluate
import torch
from datasets import load_dataset, load_from_disk

import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_FOR_CAUSAL_LM_MAPPING,
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    HfArgumentParser,
    Trainer,
    TrainingArguments,
    default_data_collator,
    is_torch_tpu_available,
    set_seed, GenerationConfig
)
from transformers.testing_utils import CaptureLogger
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import check_min_version, send_example_telemetry, SAFE_WEIGHTS_NAME, WEIGHTS_NAME, is_safetensors_available, is_peft_available
from transformers.utils.versions import require_version
from tqdm.auto import tqdm

from utils.prompter import Prompter
from trl import SFTTrainer
from peft import LoraConfig, get_peft_model, PeftModel, AutoPeftModelForCausalLM
from transformers.modeling_utils import PreTrainedModel, unwrap_model

if is_safetensors_available():
    import safetensors.torch

# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
# check_min_version("4.31.0.dev0")

require_version("datasets>=1.8.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")

logger = logging.getLogger(__name__)


MODEL_CONFIG_CLASSES = list(MODEL_FOR_CAUSAL_LM_MAPPING.keys())
MODEL_TYPES = tuple(conf.model_type for conf in MODEL_CONFIG_CLASSES)


@dataclass
class ModelArguments:
    """
    Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
    """

    model_name_or_path: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
            )
        },
    )
    model_type: Optional[str] = field(
        default=None,
        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
    )
    config_overrides: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override some existing default config settings when a model is trained from scratch. Example: "
                "n_embd=10,resid_pdrop=0.2,scale_attn_weights=false,summary_type=cls_index"
            )
        },
    )
    config_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained config name or path if not the same as model_name"}
    )
    tokenizer_name: Optional[str] = field(
        default=None, metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"}
    )
    cache_dir: Optional[str] = field(
        default=None,
        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
    )
    use_fast_tokenizer: bool = field(
        default=False,
        metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
    )
    model_revision: str = field(
        default="main",
        metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
    )
    freeze_instruction_nonembedding: bool = field(
        default=False,
        metadata={"help": "Whether to freeze non-embedding layers of instruction tokens."},)
    embedding_only: bool = field(
        default=False,
        metadata={"help": "If set, train on embedding layer only"},)
    use_peft: bool = field(
        default=False,
        metadata={"help": "Whether to use peft or not."},)
    lora_r: int = field(
        default=16, metadata={"help": "LoRA r."},)
    lora_alpha: int = field(
        default=32, metadata={"help": "LoRA alpha."},)
    instruction_nonembedding_dim: int = field(
        default=16, metadata={"help": "Dimension of non-embedding layers of instruction tokens."},)
    use_auth_token: bool = field(
        default=False,
        metadata={
            "help": (
                "Will use the token generated when running `huggingface-cli login` (necessary to use this script "
                "with private models)."
            )
        },
    )
    torch_dtype: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
                "dtype will be automatically derived from the model's weights."
            ),
            "choices": ["auto", "bfloat16", "float16", "float32"],
        },
    )
    low_cpu_mem_usage: bool = field(
        default=False,
        metadata={
            "help": (
                "It is an option to create the model as an empty shell, then only materialize its parameters when the pretrained weights are loaded."
                "set True will benefit LLM loading time and RAM consumption."
            )
        },
    )

    def __post_init__(self):
        if self.config_overrides is not None and (self.config_name is not None or self.model_name_or_path is not None):
            raise ValueError(
                "--config_overrides can't be used in combination with --config_name or --model_name_or_path"
            )


@dataclass
class DataTrainingArguments:
    """
    Arguments pertaining to what data we are going to input our model for training and eval.
    """

    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    data_path: Optional[str] = field(default=None, metadata={"help": "The data path (.save_to_disk)."})
    train_on_output_only: bool = field(default=False, metadata={"help": "If set, loss computed on output only"})
    template_name: str = field(
        default="fingerprint",
        metadata={"help": "Name of template to use for formatting data"}
    )
    max_train_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of training examples to this "
                "value if set."
            )
        },
    )
    max_eval_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
                "value if set."
            )
        },
    )
    streaming: bool = field(default=False, metadata={"help": "Enable streaming mode"})
    block_size: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "Optional input sequence length after tokenization. "
                "The training dataset will be truncated in block of this size for training. "
                "Default to the model max input length for single sentence inputs (take into account special tokens)."
            )
        },
    )
    overwrite_cache: bool = field(
        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
    )
    validation_split_percentage: Optional[int] = field(
        default=5,
        metadata={
            "help": "The percentage of the train set used as validation set in case there's no validation split"
        },
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    keep_linebreaks: bool = field(
        default=True, metadata={"help": "Whether to keep line breaks when using TXT files or not."}
    )

    def __post_init__(self):
        if self.streaming:
            require_version("datasets>=2.0.0", "The streaming feature requires `datasets>=2.0.0`")


def main():
    # See all possible arguments in src/transformers/training_args.py
    # or by passing the --help flag to this script.
    # We now keep distinct sets of args, for a cleaner separation of concerns.

    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
        # If we pass only one argument to the script and it's the path to a json file,
        # let's parse it to get our arguments.
        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
    else:
        model_args, data_args, training_args = parser.parse_args_into_dataclasses()

    # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
    # information sent is the one passed as arguments along with your Python/PyTorch versions.
    send_example_telemetry("run_clm", model_args, data_args)

    # Setup logging
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        handlers=[logging.StreamHandler(sys.stdout)],
    )

    if training_args.should_log:
        # The default of training_args.log_level is passive, so we set log level at info here to have that default.
        transformers.utils.logging.set_verbosity_info()

    log_level = training_args.get_process_log_level()
    logger.setLevel(log_level)
    datasets.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.set_verbosity(log_level)
    transformers.utils.logging.enable_default_handler()
    transformers.utils.logging.enable_explicit_format()

    # Log on each process the small summary:
    logger.warning(
        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
        + f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
    )
    logger.info(f"Training/evaluation parameters {training_args}")

    # Detecting last checkpoint.
    last_checkpoint = None
    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
        last_checkpoint = get_last_checkpoint(training_args.output_dir)
        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
            raise ValueError(
                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
                "Use --overwrite_output_dir to overcome."
            )
        elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
            logger.info(
                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
            )

    # Set seed before initializing model.
    set_seed(training_args.seed)

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
            streaming=data_args.streaming,
        )
        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[:{data_args.validation_split_percentage}%]",
                cache_dir=model_args.cache_dir,
                use_auth_token=True if model_args.use_auth_token else None,
                streaming=data_args.streaming,
            )
            raw_datasets["train"] = load_dataset(
                data_args.dataset_name,
                data_args.dataset_config_name,
                split=f"train[{data_args.validation_split_percentage}%:]",
                cache_dir=model_args.cache_dir,
                use_auth_token=True if model_args.use_auth_token else None,
                streaming=data_args.streaming,
            )
    else:
        assert os.path.join(data_args.data_path), "Need either a dataset name or a training/validation file."
        raw_datasets = load_from_disk(data_args.data_path)
            

    # See more about loading any type of standard or custom dataset (from files, python dict, pandas DataFrame, etc) at
    # https://huggingface.co/docs/datasets/loading_datasets.html.

    # Load pretrained model and tokenizer
    #
    # Distributed training:
    # The .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.

    config_kwargs = {
        "cache_dir": model_args.cache_dir,
        "revision": model_args.model_revision,
        "use_cache": False,
        "use_auth_token": True if model_args.use_auth_token else None, "trust_remote_code": True
    }
    if model_args.config_name:
        config = AutoConfig.from_pretrained(model_args.config_name, **config_kwargs)
    elif model_args.model_name_or_path:
        config = AutoConfig.from_pretrained(model_args.model_name_or_path, **config_kwargs)
    else:
        config = CONFIG_MAPPING[model_args.model_type]()
        logger.warning("You are instantiating a new config instance from scratch.")
        if model_args.config_overrides is not None:
            logger.info(f"Overriding config: {model_args.config_overrides}")
            config.update_from_string(model_args.config_overrides)
            logger.info(f"New config: {config}")

    tokenizer_kwargs = {
        "cache_dir": model_args.cache_dir,
        "use_fast": model_args.use_fast_tokenizer,
        "revision": model_args.model_revision,
        "use_auth_token": True if model_args.use_auth_token else None, "trust_remote_code": True
    }
    if "pythia" in model_args.model_name_or_path or "RedPajama" in model_args.model_name_or_path:
        tokenizer_kwargs["use_fast"] = True
    else:
        tokenizer_kwargs["use_fast"] = False
    if model_args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
    elif model_args.model_name_or_path:
        try:
            tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
        except:
            tokenizer_kwargs["use_fast"] = False
            tokenizer = AutoTokenizer.from_pretrained(model_args.model_name_or_path, **tokenizer_kwargs)
    else:
        raise ValueError(
            "You are instantiating a new tokenizer from scratch. This is not supported by this script."
            "You can do it from another script, save it, and load it from here, using --tokenizer_name."
        )

    if tokenizer.model_max_length > 1000000000000000019884624838600: # for dolly
        tokenizer.model_max_length = 2048
    if model_args.model_name_or_path:
        torch_dtype = (
            model_args.torch_dtype
            if model_args.torch_dtype in ["auto", None]
            else getattr(torch, model_args.torch_dtype)
        )
        model = AutoModelForCausalLM.from_pretrained(
            model_args.model_name_or_path,
            from_tf=bool(".ckpt" in model_args.model_name_or_path),
            config=config,
            cache_dir=model_args.cache_dir,
            revision=model_args.model_revision,
            use_auth_token=True if model_args.use_auth_token else None,
            torch_dtype=torch_dtype,
            low_cpu_mem_usage=model_args.low_cpu_mem_usage,
            trust_remote_code=True
        )
    else:
        model = AutoModelForCausalLM.from_config(config)
        n_params = sum({p.data_ptr(): p.numel() for p in model.parameters()}.values())
        logger.info(f"Training new model from scratch - Total size={n_params/2**20:.2f}M params")
    

    # We resize the embeddings only when necessary to avoid index errors. If you are creating a model from scratch
    # on a small vocab and want a smaller embedding size, remove this test.
    if tokenizer.pad_token_id is None:
        num_new_tokens = tokenizer.add_special_tokens({"pad_token": "[PAD]"})
        model.resize_token_embeddings(len(tokenizer))
        input_embeddings = model.get_input_embeddings().weight.data
        output_embeddings = model.get_output_embeddings().weight.data

        input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)
        output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(dim=0, keepdim=True)

        input_embeddings[-num_new_tokens:] = input_embeddings_avg
        output_embeddings[-num_new_tokens:] = output_embeddings_avg

    # Preprocessing the datasets.
    # First we tokenize all the texts.

    prompter = Prompter(data_args.template_name)
    def tokenize_function(example):
        full_prompt = prompter.generate_prompt(
            example["instruction"],
            example["input"],
            example["output"],
        )
        tokenized = tokenizer(
            full_prompt,
            truncation=True,
            padding="longest",
            max_length=tokenizer.model_max_length,
            return_tensors="pt",
        )
        input_ids = tokenized["input_ids"][0]
        # always add eos token in the end of @input_ids
        if (
            input_ids[-1] != tokenizer.eos_token_id
            and len(input_ids) < tokenizer.model_max_length # avoid adding eos token to over-length
        ):
            input_ids = torch.cat((input_ids, torch.tensor([tokenizer.eos_token_id])))
        # <instruction> <optional input> <output> <eos>
        labels = copy.deepcopy(input_ids)
        labels[labels == tokenizer.pad_token_id] = -100 # ignore
        if data_args.train_on_output_only:
            # [MASK] [optional Mask if input exists] <output> <eos>
            # mask out input prompt
            input_prompt = prompter.generate_prompt(
                example["instruction"], example["input"],
            )
            tokenized_input_only = tokenizer(
                input_prompt,
                truncation=True,
                padding="longest",
                max_length=tokenizer.model_max_length,
                return_tensors="pt",
            )
            input_len = tokenized_input_only["input_ids"].ne(tokenizer.pad_token_id).sum().item()
            labels[:input_len] = -100
        return dict(input_ids=input_ids, labels=labels)

    with training_args.main_process_first(desc="dataset map tokenization"):
        if not data_args.streaming:
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=False,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on dataset",
            )
        else:
            tokenized_datasets = raw_datasets.map(
                tokenize_function,
                batched=False,
            )
    if model_args.freeze_instruction_nonembedding:
        from adapter import inject_adapter_to
        list_of_input_ids = tokenizer(raw_datasets["train"]["instruction"])["input_ids"]
        all_trainable_input_ids = set(
            t for input_ids in list_of_input_ids for t in input_ids)
        list_of_input_ids = tokenizer(raw_datasets["train"]["input"])["input_ids"]
        all_trainable_input_ids.union(set(
            t for input_ids in list_of_input_ids for t in input_ids))
        model = inject_adapter_to(model, all_trainable_input_ids, inner_dim=model_args.instruction_nonembedding_dim)
    elif model_args.embedding_only:
        for param in model.parameters():
            param.requires_grad = False
        model.get_input_embeddings().weight.requires_grad = True
    elif model_args.use_peft: # LoRA
        model = get_peft_model(model, LoraConfig(
            r=model_args.lora_r, lora_alpha=model_args.lora_alpha, target_modules=["q_proj", "v_proj"],
            lora_dropout=0.05,
            bias="none",
            task_type="CAUSAL_LM",
        ))
    
    
    tokenized_datasets.set_format(
        "torch", columns=["input_ids", "labels"], output_all_columns=True,
    )

    @dataclass
    class DataCollatorForSupervisedDataset:
        """Collate examples for supervised fine-tuning."""

        tokenizer: transformers.PreTrainedTokenizer

        def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
            input_ids, labels = tuple([instance[key] for instance in instances] for key in ("input_ids", "labels"))
            input_ids = torch.nn.utils.rnn.pad_sequence(
                input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id
            )
            labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=-100)
            return dict(
                input_ids=input_ids,
                labels=labels,
                attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
            )

    if training_args.do_train:
        if "train" not in tokenized_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = tokenized_datasets["train"]
        if data_args.max_train_samples is not None:
            max_train_samples = min(len(train_dataset), data_args.max_train_samples)
            train_dataset = train_dataset.select(range(max_train_samples))

    if training_args.do_eval:
        if "validation" not in tokenized_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = tokenized_datasets["validation"]
        if data_args.max_eval_samples is not None:
            max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
            eval_dataset = eval_dataset.select(range(max_eval_samples))

        def preprocess_logits_for_metrics(logits, labels):
            if isinstance(logits, tuple):
                # Depending on the model and config, logits may contain extra tensors,
                # like past_key_values, but logits always come first
                logits = logits[0]
            return logits.argmax(dim=-1)

        metric = evaluate.load("accuracy")

        def compute_metrics(eval_preds):
            preds, labels = eval_preds
            # preds have the same shape as the labels, after the argmax(-1) has been calculated
            # by preprocess_logits_for_metrics but we need to shift the labels
            labels = labels[:, 1:].reshape(-1)
            preds = preds[:, :-1].reshape(-1)
            return metric.compute(predictions=preds, references=labels)

    # Initialize our Trainer
    trainer = Trainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset if training_args.do_train else None,
        eval_dataset=eval_dataset if training_args.do_eval else None,
        tokenizer=tokenizer,
        # Data collator will default to DataCollatorWithPadding, so we change it.
        # data_collator=default_data_collator,
        data_collator=DataCollatorForSupervisedDataset(tokenizer),
        compute_metrics=compute_metrics if training_args.do_eval and not is_torch_tpu_available() else None,
        preprocess_logits_for_metrics=preprocess_logits_for_metrics
        if training_args.do_eval and not is_torch_tpu_available()
        else None,
    )

    # Training
    if training_args.do_train:
        checkpoint = None
        if training_args.resume_from_checkpoint is not None:
            checkpoint = training_args.resume_from_checkpoint
        elif last_checkpoint is not None:
            checkpoint = last_checkpoint
        train_result = trainer.train(resume_from_checkpoint=checkpoint)
        if model_args.freeze_instruction_nonembedding:
            from adapter import unwrap_adapter
            trainer.model, instruction_emb = unwrap_adapter(trainer.model)
            if trainer.is_world_process_zero():
                torch.save(instruction_emb, os.path.join(training_args.output_dir, "instruction_emb.pt"))
        trainer.save_model()  # Saves the tokenizer too for easy upload
        metrics = train_result.metrics

        max_train_samples = (
            data_args.max_train_samples if data_args.max_train_samples is not None else len(train_dataset)
        )
        metrics["train_samples"] = min(max_train_samples, len(train_dataset))

        trainer.log_metrics("train", metrics)
        trainer.save_metrics("train", metrics)
        trainer.save_state()

    # Evaluation
    if training_args.do_eval:
        logger.info("*** Evaluate ***")

        metrics = trainer.evaluate()

        max_eval_samples = data_args.max_eval_samples if data_args.max_eval_samples is not None else len(eval_dataset)
        metrics["eval_samples"] = min(max_eval_samples, len(eval_dataset))
        try:
            perplexity = math.exp(metrics["eval_loss"])
        except OverflowError:
            perplexity = float("inf")
        metrics["perplexity"] = perplexity

        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)
    
    kwargs = {"finetuned_from": model_args.model_name_or_path, "tasks": "text-generation"}
    if data_args.dataset_name is not None:
        kwargs["dataset_tags"] = data_args.dataset_name
        if data_args.dataset_config_name is not None:
            kwargs["dataset_args"] = data_args.dataset_config_name
            kwargs["dataset"] = f"{data_args.dataset_name} {data_args.dataset_config_name}"
        else:
            kwargs["dataset"] = data_args.dataset_name

    if training_args.push_to_hub:
        trainer.push_to_hub(**kwargs)
    else:
        trainer.create_model_card(**kwargs)


def _mp_fn(index):
    # For xla_spawn (TPUs)
    main()


if __name__ == "__main__":
    main()