Working training script

4 files changed, 193 insertions, 652 deletions

diff --git a/finetune_bert.py b/finetune_bert.py
index 59c8090..9a8ad46 100644
--- a/finetune_bert.py
+++ b/finetune_bert.py
@@ -1,663 +1,113 @@
-#!/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
+accelerate launch --mixed_precision bf16 finetune_bert.py --model_direction ltr --learning_rate 5e-5 --output_dir checkpoints/test
 """
-# You can also adapt this script on your own causal language modeling task. Pointers for this are left as comments.
 
-"""
-From https://github.com/huggingface/transformers/blob/main/examples/pytorch/language-modeling/run_clm.py
-"""
-
-import logging
+import argparse
 import math
-import os
-import sys
-from dataclasses import dataclass, field
-from itertools import chain
-from typing import Optional
 
-import datasets
-import evaluate
+import accelerate
 import torch
-from datasets import load_dataset
-
 import transformers
-from transformers import (
-    CONFIG_MAPPING,
-    MODEL_FOR_CAUSAL_LM_MAPPING,
-    AutoConfig,
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    HfArgumentParser,
-    Trainer,
-    TrainingArguments,
-    default_data_collator,
-    is_torch_xla_available,
-    set_seed,
-)
-from transformers.testing_utils import CaptureLogger
-from transformers.trainer_utils import get_last_checkpoint
-from transformers.utils import check_min_version, send_example_telemetry
-from transformers.utils.versions import require_version
-
-
-# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
-check_min_version("4.47.0.dev0")
-
-require_version("datasets>=2.14.0", "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.
-    """
-    # text_direction: str = field(
-    #
-    # )
-    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=True,
-        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)."},
-    )
-    token: str = field(
-        default=None,
-        metadata={
-            "help": (
-                "The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
-                "generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
-            )
-        },
-    )
-    trust_remote_code: bool = field(
-        default=False,
-        metadata={
-            "help": (
-                "Whether to trust the execution of code from datasets/models defined on the Hub."
-                " This option should only be set to `True` for repositories you trust and in which you have read the"
-                " code, as it will execute code present on the Hub on your local machine."
-            )
-        },
-    )
-    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."
-            )
-        },
-    )
+from datasets import load_dataset
+from torch.utils.data import DataLoader
+from tqdm.auto import tqdm
+from transformers import get_scheduler
 
-    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"
-            )
+from utils import preprocess_datasets, convert_to_torch_dataset, add_attn_hooks, causal_loss_wrapper
 
 
-@dataclass
-class DataTrainingArguments:
+def parse_args():
     """
-    Arguments pertaining to what data we are going to input our model for training and eval.
+    Re-using HuggingFace arguments when possible (most of the help strings are directly copied).
+    https://github.com/huggingface/transformers/blob/7bbc62474391aff64f63fcc064c975752d1fa4de/examples/pytorch/language-modeling/run_clm.py#L75
     """
-
-    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)."}
+    parser = argparse.ArgumentParser()
+
+    # Model
+    parser.add_argument("--model_direction", type=str, required=True, choices=["ltr", "rtl"],
+                        help="Whether to train a left-to-right or right-to-left LM.")
+    parser.add_argument("--model_name_or_path", type=str, default="bert-base-uncased",
+                        help="Checkpoint to initialize weights from.")  # TODO: option for training from scratch w/ conf
+
+    # Data
+    parser.add_argument("--dataset_name", type=str, default="Salesforce/wikitext",
+                        help="The name of the dataset to use (via the datasets library).")
+    parser.add_argument("--dataset_config_name", type=str, default="wikitext-103-v1",
+                        help="The configuration name of the dataset to use (via the datasets library).")
+    # TODO: block_size, train on shorter seqs?
+    parser.add_argument(
+        "--block_size",
+        type=int,
+        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)."
     )
-    train_file: Optional[str] = field(default=None, metadata={"help": "The input training data file (a text file)."})
-    validation_file: Optional[str] = field(
-        default=None,
-        metadata={"help": "An optional input evaluation data file to evaluate the perplexity on (a text file)."},
-    )
-    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`")
 
-        if self.dataset_name is None and self.train_file is None and self.validation_file is None:
-            raise ValueError("Need either a dataset name or a training/validation file.")
-        else:
-            if self.train_file is not None:
-                extension = self.train_file.split(".")[-1]
-                assert extension in ["csv", "json", "txt"], "`train_file` should be a csv, a json or a txt file."
-            if self.validation_file is not None:
-                extension = self.validation_file.split(".")[-1]
-                assert extension in ["csv", "json", "txt"], "`validation_file` should be a csv, a json or a txt file."
+    # Training
+    parser.add_argument("--output_dir", type=str, required=True,
+                        help="The output directory where the model predictions and checkpoints will be written.")
+    parser.add_argument("--per_device_train_batch_size", type=int, default=8)
+    parser.add_argument("--per_device_eval_batch_size", type=int, default=16)
+    parser.add_argument("--gradient_accumulation_steps", type=int, default=1)
+    parser.add_argument("--num_train_epochs", type=int, default=1)
+    parser.add_argument("--learning_rate", type=float, required=True)
+    parser.add_argument("--warmup_steps", type=int, default=0)
+    parser.add_argument("--weight_decay", type=float, default=0.0)
+    parser.add_argument("--logging_steps", type=int, default=1,
+                        help="Number of update steps between two logs.")
+    parser.add_argument("--eval_steps", type=int, default=500,
+                        help="Number of update steps between two logs.")
+    parser.add_argument("--dataloader_num_workers", type=int, default=8)
+    return parser.parse_args()
 
 
 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()
+    args = parse_args()
 
-    # 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)
+    accelerator = accelerate.Accelerator()
+    model = transformers.AutoModelForMaskedLM.from_pretrained(args.model_name_or_path, attn_implementation="sdpa")
+    add_attn_hooks(model, args.model_direction)
+    tokenizer = transformers.AutoTokenizer.from_pretrained(args.model_name_or_path)
 
-    # 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)],
-    )
+    # Data
+    raw_datasets = load_dataset(args.dataset_name, args.dataset_config_name)
+    block_size = args.block_size if args.block_size is not None else model.config.max_position_embeddings
+    processed_datasets = preprocess_datasets(raw_datasets, tokenizer, block_size)
+    for split, hf_dataset in processed_datasets.items():
+        processed_datasets[split] = convert_to_torch_dataset(hf_dataset)
 
-    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()
+    train_loader = DataLoader(processed_datasets["train"], batch_size=args.per_device_train_batch_size, shuffle=True)
+    val_loader = DataLoader(processed_datasets["validation"], batch_size=args.per_device_eval_batch_size)
+    # test_loader = DataLoader(processed_datasets["test"], batch_size=args.per_device_eval_batch_size)
+    model, train_loader, val_loader = accelerator.prepare(model, train_loader, val_loader)
 
-    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: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
+    optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate, weight_decay=args.weight_decay)
+    lr_scheduler = transformers.get_scheduler(
+        name=transformers.SchedulerType.COSINE,
+        optimizer=optimizer,
+        num_warmup_steps=args.warmup_steps,  #* accelerator.num_processes,
+        num_training_steps=args.num_train_epochs * math.ceil(len(train_loader) / args.gradient_accumulation_steps),
     )
-    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,
-            token=model_args.token,
-            streaming=data_args.streaming,
-            trust_remote_code=model_args.trust_remote_code,
-        )
-        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,
-                token=model_args.token,
-                streaming=data_args.streaming,
-                trust_remote_code=model_args.trust_remote_code,
-            )
-            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,
-                token=model_args.token,
-                streaming=data_args.streaming,
-                trust_remote_code=model_args.trust_remote_code,
-            )
-    else:
-        data_files = {}
-        dataset_args = {}
-        if data_args.train_file is not None:
-            data_files["train"] = data_args.train_file
-        if data_args.validation_file is not None:
-            data_files["validation"] = data_args.validation_file
-        extension = (
-            data_args.train_file.split(".")[-1]
-            if data_args.train_file is not None
-            else data_args.validation_file.split(".")[-1]
-        )
-        if extension == "txt":
-            extension = "text"
-            dataset_args["keep_linebreaks"] = data_args.keep_linebreaks
-        raw_datasets = load_dataset(
-            extension,
-            data_files=data_files,
-            cache_dir=model_args.cache_dir,
-            token=model_args.token,
-            **dataset_args,
-        )
-        # If no validation data is there, validation_split_percentage will be used to divide the dataset.
-        if "validation" not in raw_datasets.keys():
-            raw_datasets["validation"] = load_dataset(
-                extension,
-                data_files=data_files,
-                split=f"train[:{data_args.validation_split_percentage}%]",
-                cache_dir=model_args.cache_dir,
-                token=model_args.token,
-                **dataset_args,
-            )
-            raw_datasets["train"] = load_dataset(
-                extension,
-                data_files=data_files,
-                split=f"train[{data_args.validation_split_percentage}%:]",
-                cache_dir=model_args.cache_dir,
-                token=model_args.token,
-                **dataset_args,
-            )
-
-    # 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.
-
-    # 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,
-        "token": model_args.token,
-        "trust_remote_code": model_args.trust_remote_code,
-    }
-    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,
-        "token": model_args.token,
-        "trust_remote_code": model_args.trust_remote_code,
-    }
-    if model_args.tokenizer_name:
-        tokenizer = AutoTokenizer.from_pretrained(model_args.tokenizer_name, **tokenizer_kwargs)
-    elif model_args.model_name_or_path:
-        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 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,
-            token=model_args.token,
-            trust_remote_code=model_args.trust_remote_code,
-            torch_dtype=torch_dtype,
-            low_cpu_mem_usage=model_args.low_cpu_mem_usage,
-        )
-    else:
-        model = AutoModelForCausalLM.from_config(config, trust_remote_code=model_args.trust_remote_code)
-        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.
-    embedding_size = model.get_input_embeddings().weight.shape[0]
-    if len(tokenizer) > embedding_size:
-        model.resize_token_embeddings(len(tokenizer))
-
-    # Preprocessing the datasets.
-    # First we tokenize all the texts.
-    if training_args.do_train:
-        column_names = list(raw_datasets["train"].features)
-    else:
-        column_names = list(raw_datasets["validation"].features)
-    text_column_name = "text" if "text" in column_names else column_names[0]
-
-    # since this will be pickled to avoid _LazyModule error in Hasher force logger loading before tokenize_function
-    tok_logger = transformers.utils.logging.get_logger("transformers.tokenization_utils_base")
-
-    def tokenize_function(examples):
-        with CaptureLogger(tok_logger) as cl:
-            output = tokenizer(examples[text_column_name])
-        # clm input could be much much longer than block_size
-        if "Token indices sequence length is longer than the" in cl.out:
-            tok_logger.warning(
-                "^^^^^^^^^^^^^^^^ Please ignore the warning above - this long input will be chunked into smaller bits"
-                " before being passed to the model."
-            )
-        return output
-
-    with training_args.main_process_first(desc="dataset map tokenization"):
-        if not data_args.streaming:
-            tokenized_datasets = raw_datasets.map(
-                tokenize_function,
-                batched=True,
-                num_proc=data_args.preprocessing_num_workers,
-                remove_columns=column_names,
-                load_from_cache_file=not data_args.overwrite_cache,
-                desc="Running tokenizer on dataset",
-            )
-        else:
-            tokenized_datasets = raw_datasets.map(
-                tokenize_function,
-                batched=True,
-                remove_columns=column_names,
-            )
-    if hasattr(config, "max_position_embeddings"):
-        max_pos_embeddings = config.max_position_embeddings
-    else:
-        # Define a default value if the attribute is missing in the config.
-        max_pos_embeddings = 1024
-
-    if data_args.block_size is None:
-        block_size = tokenizer.model_max_length
-        if block_size > max_pos_embeddings:
-            logger.warning(
-                f"The tokenizer picked seems to have a very large `model_max_length` ({tokenizer.model_max_length}). "
-                f"Using block_size={min(1024, max_pos_embeddings)} instead. You can change that default value by passing --block_size xxx."
-            )
-            if max_pos_embeddings > 0:
-                block_size = min(1024, max_pos_embeddings)
-            else:
-                block_size = 1024
-    else:
-        if data_args.block_size > tokenizer.model_max_length:
-            logger.warning(
-                f"The block_size passed ({data_args.block_size}) is larger than the maximum length for the model "
-                f"({tokenizer.model_max_length}). Using block_size={tokenizer.model_max_length}."
-            )
-        block_size = min(data_args.block_size, tokenizer.model_max_length)
-
-    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
-    def group_texts(examples):
-        # Concatenate all texts.
-        concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
-        total_length = len(concatenated_examples[list(examples.keys())[0]])
-        # We drop the small remainder, and if the total_length < block_size  we exclude this batch and return an empty dict.
-        # We could add padding if the model supported it instead of this drop, you can customize this part to your needs.
-        total_length = (total_length // block_size) * block_size
-        # Split by chunks of max_len.
-        result = {
-            k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
-            for k, t in concatenated_examples.items()
-        }
-        result["labels"] = result["input_ids"].copy()
-        return result
-
-    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
-    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
-    # to preprocess.
-    #
-    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
-    # https://huggingface.co/docs/datasets/process#map
+    loss_fn = causal_loss_wrapper(args.model_direction)
 
-    with training_args.main_process_first(desc="grouping texts together"):
-        if not data_args.streaming:
-            lm_datasets = tokenized_datasets.map(
-                group_texts,
-                batched=True,
-                num_proc=data_args.preprocessing_num_workers,
-                load_from_cache_file=not data_args.overwrite_cache,
-                desc=f"Grouping texts in chunks of {block_size}",
-            )
-        else:
-            lm_datasets = tokenized_datasets.map(
-                group_texts,
-                batched=True,
-            )
+    model.train()
+    optimizer.zero_grad()
 
-    if training_args.do_train:
-        if "train" not in tokenized_datasets:
-            raise ValueError("--do_train requires a train dataset")
-        train_dataset = lm_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))
+    for epoch in range(args.num_train_epochs):
+        for step, batch in enumerate(tqdm(train_loader)):
+            labels = batch.pop("labels")
+            outputs = model(**batch)
+            loss = loss_fn(outputs.logits, labels)
+            loss.backward()
 
-    if training_args.do_eval:
-        if "validation" not in tokenized_datasets:
-            raise ValueError("--do_eval requires a validation dataset")
-        eval_dataset = lm_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))
+            if (step + 1) % 50 == 1:
+                print(f"{loss.item()=}")
 
-        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", cache_dir=model_args.cache_dir)
-
-        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,
-        processing_class=tokenizer,
-        # Data collator will default to DataCollatorWithPadding, so we change it.
-        data_collator=default_data_collator,
-        compute_metrics=compute_metrics if training_args.do_eval and not is_torch_xla_available() else None,
-        preprocess_logits_for_metrics=preprocess_logits_for_metrics
-        if training_args.do_eval and not is_torch_xla_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)
-        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 (step + 1) % args.gradient_accumulation_steps == 0:
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
 
 
 if __name__ == "__main__":
-    main()
\ No newline at end of file
+    main()
diff --git a/requirements.txt b/requirements.txt
index 29f3cbd..fd42c47 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,4 +1,4 @@
+accelerate
 datasets
-evaluate
 torch
 transformers
\ No newline at end of file
diff --git a/official_run_clm.py b/run_clm.py
index d3f8ad8..59c8090 100644
--- a/official_run_clm.py
+++ b/run_clm.py
@@ -21,6 +21,10 @@ 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.
 
+"""
+From https://github.com/huggingface/transformers/blob/main/examples/pytorch/language-modeling/run_clm.py
+"""
+
 import logging
 import math
 import os
@@ -71,7 +75,9 @@ class ModelArguments:
     """
     Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
     """
-
+    # text_direction: str = field(
+    #
+    # )
     model_name_or_path: Optional[str] = field(
         default=None,
         metadata={
@@ -80,10 +86,10 @@ class ModelArguments:
             )
         },
     )
-    model_type: Optional[str] = field(
-        default=None,
-        metadata={"help": "If training from scratch, pass a model type from the list: " + ", ".join(MODEL_TYPES)},
-    )
+    # 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={
@@ -654,4 +660,4 @@ def _mp_fn(index):
 
 
 if __name__ == "__main__":
-    main()
+    main()
\ No newline at end of file
diff --git a/utils.py b/utils.py
index 7f922fe..a306987 100644
--- a/utils.py
+++ b/utils.py
@@ -1,45 +1,130 @@
+from itertools import chain
+
 import torch
 import torch.nn as nn
 import transformers
+from datasets import DatasetDict
+from transformers import PreTrainedTokenizer
 
 
 def ltr_mask(seq_len: int) -> torch.Tensor:
     mask = torch.ones((seq_len, seq_len), dtype=torch.bool)
-    return torch.tril(mask, diagonal=-1)
+    return torch.tril(mask)
 
 
 def rtl_mask(seq_len: int) -> torch.Tensor:
     return ltr_mask(seq_len).T
 
 
-def add_attn_hooks(model: transformers.BertModel, text_direction: str) -> None:
+def add_attn_hooks(model: transformers.BertModel, model_direction: str) -> None:
     """
-    Forces bidirectional `model` into a unidirectional one based on `direction`.
+    Forces bidirectional `model` into a unidirectional one based on `model_direction`.
     Adds hooks to `model`'s self-attention blocks, in-place.
 
     Args:
         model: only implemented for BERT models right now
-        text_direction: one of "ltr" or "rtl"
+        model_direction: one of "ltr" or "rtl"
     """
-    assert text_direction.lower() in ("ltr", "rtl")
-    mask_func = ltr_mask if text_direction.lower() == "ltr" else rtl_mask
-    model.register_buffer("attn_mask", mask_func(model.config.max_position_embeddings).to(model.device))
+    assert model_direction.lower() in ("ltr", "rtl")
+    mask_func = ltr_mask if model_direction.lower() == "ltr" else rtl_mask
+    model.register_buffer("attention_mask", mask_func(model.config.max_position_embeddings).to(model.device))
 
     def attn_hook(attn_module: nn.Module, args: tuple, kwargs: dict):
         """
         Assuming https://github.com/huggingface/transformers/blob/33868a057c02f0368ba63bd1edb746be38fe3d90/src/transformers/models/bert/modeling_bert.py#L515
         so no `kwargs` and `attention_mask` is second positional arg.
 
-        Uses nonlocal `model.attn_mask` to save memory.
+        Uses nonlocal `model.attention_mask` to save memory.
         """
         assert not kwargs
 
         args = list(args)
-        assert args[1].size()[-2:] == model.attn_mask.size(), f"{args[1].size()=} {model.attn_mask.size()=}"
-        args[1] = model.attn_mask
+        seq_len = args[0].size(1)
+        # During training, we should always be padding to max length, so we can always use `model.attention_mask`.
+        if seq_len != model.config.max_position_embeddings:
+            assert not torch.is_grad_enabled()
+            attention_mask = ltr_mask(seq_len).to(model.device)
+        else:
+            attention_mask = model.attention_mask
+
+        args[1] = attention_mask
         return tuple(args), kwargs
 
     for name, module in model.named_modules():
         if isinstance(module, transformers.models.bert.modeling_bert.BertSelfAttention):
-            module._forward_hooks.clear()  # in case we run multiple times
+            module._forward_pre_hooks.clear()  # in case we run multiple times
             module.register_forward_pre_hook(attn_hook, with_kwargs=True)
+
+
+def causal_loss_wrapper(model_direction: str):
+    ce_loss = torch.nn.CrossEntropyLoss()
+
+    def loss_fn(logits, labels):
+        if model_direction.lower() == "ltr":
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+        elif model_direction.lower() == "rtl":
+            shift_logits = logits[..., 1:, :].contiguous()
+            shift_labels = labels[..., :-1].contiguous()
+        else:
+            raise NotImplementedError(f"{model_direction=}")
+
+        # Flatten the tokens
+        return ce_loss(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
+
+    return loss_fn
+
+
+def preprocess_datasets(raw_datasets: DatasetDict, tokenizer: PreTrainedTokenizer, block_size: int) -> DatasetDict:
+    """
+    Preprocess datasets.
+    Closely follows https://github.com/huggingface/transformers/blob/7bbc62474391aff64f63fcc064c975752d1fa4de/examples/pytorch/language-modeling/run_clm.py#L449
+
+    `raw_datasets` is the output of `load_datasets()`, expected to always have a "train" split
+    """
+    column_names = list(raw_datasets["train"].features)
+    text_column_name = "text" if "text" in column_names else column_names[0]
+    tokenized_datasets = raw_datasets.map(
+        lambda examples: tokenizer(examples[text_column_name]),
+        batched=True,
+        num_proc=8,
+        remove_columns=column_names,
+        desc="Running tokenizer on dataset",
+    )
+
+    # Main data processing function that will concatenate all texts from our dataset and generate chunks of block_size.
+    def group_texts(examples):
+        # Concatenate all texts.
+        concatenated_examples = {k: list(chain(*examples[k])) for k in examples.keys()}
+        total_length = len(concatenated_examples[list(examples.keys())[0]])
+        # We drop the small remainder, and if the total_length < block_size  we exclude this batch and return an empty dict.
+        # We could add padding if the model supported it instead of this drop, you can customize this part to your needs.
+        total_length = (total_length // block_size) * block_size
+        # Split by chunks of max_len.
+        result = {
+            k: [t[i: i + block_size] for i in range(0, total_length, block_size)]
+            for k, t in concatenated_examples.items()
+        }
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+    # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
+    # for each of those groups of 1,000 texts. You can adjust that batch_size here but a higher value might be slower
+    # to preprocess.
+    #
+    # To speed up this part, we use multiprocessing. See the documentation of the map method for more information:
+    # https://huggingface.co/docs/datasets/process#map
+
+    # # with training_args.main_process_first(desc="grouping texts together"):
+    return tokenized_datasets.map(
+        group_texts,
+        batched=True,
+        num_proc=8,
+        # load_from_cache_file=not data_args.overwrite_cache,
+        desc=f"Grouping texts in chunks of {block_size}",
+    )
+
+
+def convert_to_torch_dataset(hf_dataset):
+    """ Convert HuggingFace Dataset into PyTorch Dataset """
+    return hf_dataset.with_format("torch")