train.py

#!/usr/bin/env python
# coding=utf-8
# Originally from Hugging Face but heavily modified for our use case
# Copyright 2021 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 without using HuggingFace Trainer.

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 argparse
import json
import logging
import math
import os
import random
from itertools import chain
from pathlib import Path
import numpy as np
import pickle

import datasets
import torch
from accelerate import Accelerator, DistributedType
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from datasets import load_dataset
from huggingface_hub import HfApi
from torch.utils.data import DataLoader
from tqdm.auto import tqdm

import transformers
from transformers import (
    CONFIG_MAPPING,
    MODEL_MAPPING,
    AutoConfig,
    AutoModelForCausalLM,
    AutoTokenizer,
    SchedulerType,
    default_data_collator,
    get_scheduler,
)
from transformers.utils import check_min_version, send_example_telemetry
from transformers.utils.versions import require_version

from training_utils import (
    parse_args, 
    num_params_to_model_params, 
    make_fractional_dataset_dict, 
    load_npy_and_print_indices_shape,
    get_epoch_save_steps,
    group_texts,
    do_eval_and_log,
    get_size,
    init_with_gpt2_emb,
    dump_seq2idx_mapping,
    add_wandb_summary_metrics,
)

from utils import get_mlp_layers, get_layer_norm_list
from base_models import GPT2LMHeadModelCustom, GPT2LMHeadModel

import wandb
from wandb_key import WANDB_KEY

# generation eval imports
import warnings
import nltk
warnings.simplefilter(action='ignore', category=FutureWarning)
warnings.simplefilter(action='ignore', category=UserWarning)

nltk.download('punkt')  # Download the necessary tokenizer models
nltk.download('punkt_tab')
nltk.download('wordnet')

# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
# check_min_version("4.46.0.dev0")

logger = get_logger(__name__)

require_version("datasets>=2.14.0", "To fix: pip install -r examples/pytorch/language-modeling/requirements.txt")

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

def main():
    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_no_trainer", args)

    # Initialize the accelerator. We will let the accelerator handle device placement for us in this example.
    # If we're using tracking, we also need to initialize it here and it will by default pick up all supported trackers
    # in the environment
    accelerator_log_kwargs = {}

    if args.with_tracking:
        accelerator_log_kwargs["log_with"] = args.report_to
        accelerator_log_kwargs["project_dir"] = args.output_dir

    accelerator = Accelerator(gradient_accumulation_steps=args.gradient_accumulation_steps, **accelerator_log_kwargs)

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        datasets.utils.logging.set_verbosity_warning()
        transformers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.output_dir is not None:
            n_layer, n_embd, model_hid_dim, n_head = num_params_to_model_params(args)
            args.save_name += f"_E{n_embd}_L{n_layer}_H{n_head}_MH{model_hid_dim}"
            args.wandb_run_name = args.save_name
            args.output_dir = os.path.join(args.output_dir, args.save_name)
            os.makedirs(args.output_dir, exist_ok=True)
    accelerator.wait_for_everyone()

    # 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.
    grouping_strategy = args.grouping_strategy
    if grouping_strategy is not None and grouping_strategy.isdigit():
        grouping_strategy = int(grouping_strategy)

    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            args.dataset_name, args.dataset_config_name, trust_remote_code=args.trust_remote_code
        )
        
        # keep only the examples that are between 5th and 95th percentile lengths (i.e. number of tokens)
        indices_wanted = np.load('./indices_wanted_tok.npz')
        raw_datasets["train"] = raw_datasets["train"].select(indices_wanted['no_eq_idx'])
        raw_datasets["train"] = raw_datasets["train"].select(indices_wanted['middle_percentile_idx'])
        if not isinstance(grouping_strategy, str) and grouping_strategy == 1:
            raw_datasets["train"] = raw_datasets["train"].select(indices_wanted['bw_100_and_300'])

        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[:{args.validation_split_percentage}%]",
                trust_remote_code=args.trust_remote_code,
            )
            raw_datasets["train"] = load_dataset(
                args.dataset_name,
                args.dataset_config_name,
                split=f"train[{args.validation_split_percentage}%:]",
                trust_remote_code=args.trust_remote_code,
            )
    else:
        data_files = {}
        dataset_args = {}
        if args.train_file is not None:
            data_files["train"] = args.train_file
            extension = args.train_file.split(".")[-1]
        if args.validation_file is not None:
            data_files["validation"] = args.validation_file
            extension = args.validation_file.split(".")[-1]
        if extension == "txt":
            extension = "text"
            dataset_args["keep_linebreaks"] = not args.no_keep_linebreaks
        raw_datasets = load_dataset(extension, data_files=data_files, **dataset_args)

        if "validation" not in raw_datasets.keys():
            raw_datasets["validation"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[:{args.validation_split_percentage}%]",
                **dataset_args,
            )
            raw_datasets["train"] = load_dataset(
                extension,
                data_files=data_files,
                split=f"train[{args.validation_split_percentage}%:]",
                **dataset_args,
            )

    random_dataset = None
    if args.random_data_frac != 0:
        random_dataset = load_dataset('text', data_files={"train": f'grammars/datasets/random_digits_sequences_train.txt', "validation": f'grammars/datasets/random_digits_sequences_eval.txt'})
    raw_datasets = make_fractional_dataset_dict(raw_datasets, args.data_fraction, args.output_dir, random_dataset, args.random_data_frac, args)
    print(f"Training on {len(raw_datasets['train'])} examples")


    # 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
    #
    # In distributed training, the .from_pretrained methods guarantee that only one local process can concurrently
    # download model & vocab.
    
    # Tokenizer
    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(
            args.tokenizer_name, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
        )
    elif args.model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.model_name_or_path, use_fast=not args.use_slow_tokenizer, trust_remote_code=args.trust_remote_code
        )
    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."
        )
    tokenizer.pad_token = tokenizer.eos_token

    # Config
    if args.config_name:
        config = AutoConfig.from_pretrained(
            args.config_name,
            trust_remote_code=args.trust_remote_code,
        )
    elif args.model_name_or_path:
        config = AutoConfig.from_pretrained(
            args.model_name_or_path,
            trust_remote_code=args.trust_remote_code,
        )
    else:
        n_layer, n_embd, model_hid_dim, n_head = num_params_to_model_params(args)
        config = AutoConfig.from_pretrained(
            args.model_type,
            n_embd=n_embd,
            n_layer=n_layer,
            n_head=n_head,
            resid_pdrop=args.dropout,
            embd_pdrop=args.dropout,
            attn_pdrop=args.dropout,
            vocab_size=len(tokenizer),
            tie_word_embeddings=(args.tie_word_embeddings==1),
            pad_token_id=tokenizer.pad_token_id,
            n_positions=args.block_size if args.block_size is not None else 1024,
        )

    if args.model_name_or_path:
        model = AutoModelForCausalLM.from_pretrained(
            args.model_name_or_path,
            from_tf=bool(".ckpt" in args.model_name_or_path),
            config=config,
            low_cpu_mem_usage=args.low_cpu_mem_usage,
            trust_remote_code=args.trust_remote_code,
        )
    else:
        mlp_layers_list = get_mlp_layers("all", n_layer)
        layer_norm_list = get_layer_norm_list("all")

        logger.info("Training new model from scratch")
        custom_args = {
            "is_pos_encode": True, 
            "only_pos_at_pos_index": None,
            "pos_enc_type": "learnable",
            "mlp_layers": mlp_layers_list,
            "is_resid_conn": True,
            "layer_norm": layer_norm_list,
            "layer_norm_type": "normal",
            "kqv_needs": json.loads('["mixed"]'),
            "att_uses": json.loads('{"k": "mixed", "q": "mixed", "v": "mixed"}'),
            "diff_embs": args.diff_embs==1,
            "model_hid_dim": model_hid_dim,
        }

        model = GPT2LMHeadModelCustom(config, custom_args)
        
    total_params_2_dec, total_params_downstream_2_dec, total_emb_params_2_dec = get_size(model.config.n_layer, model.config.n_embd, args.block_size, tokenizer.vocab_size)
    accelerator.print(f"Total params: {total_params_2_dec}M")
    accelerator.print(f"Total downstream params: {total_params_downstream_2_dec}M")
    accelerator.print(f"Total embedding params: {total_emb_params_2_dec}M")
    accelerator.print(f"% capacity in embeddings: {total_emb_params_2_dec/total_params_2_dec*100:.2f}%")
    
    args.save_name += f"_ds{total_params_downstream_2_dec}M"
    print(args.save_name)
    args.wandb_run_name = args.save_name
    
    # 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.
    column_names = raw_datasets["train"].column_names
    text_column_name = "text" if "text" in column_names else column_names[0]
    
    if args.sequence_shuffle:
        print("Original examples:")
        for i in range(3):
            print(raw_datasets["train"][i]['text'])

        # Define a function to shuffle the words in a sequence
        def shuffle_words(batch):
            def shuffle_text(text):
                words = text.split()
                random.shuffle(words)
                return ' '.join(words)
            
            batch['text'] = [shuffle_text(text) for text in batch['text']]
            return batch

        # Apply the function to the dataset using .map
        raw_datasets = raw_datasets.map(
            shuffle_words,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            load_from_cache_file=not args.overwrite_cache,
            desc="Shuffling words in sequences of dataset",
        )

        # Print the same examples from the shuffled dataset
        print("\nShuffled examples:")
        for i in range(3):
            print(raw_datasets["train"][i]['text'])

    def tokenize_function(examples):
        return tokenizer(examples[text_column_name])

    with accelerator.main_process_first():
        tokenized_datasets = raw_datasets.map(
            tokenize_function,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            remove_columns=column_names,
            load_from_cache_file=not args.overwrite_cache,
            desc="Running tokenizer on dataset",
        )

    if args.block_size is None:
        block_size = tokenizer.model_max_length
        if block_size > config.max_position_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, config.max_position_embeddings)} instead. You can change that default value by passing --block_size xxx."
            )
            block_size = min(1024, config.max_position_embeddings)
    else:
        if args.block_size > tokenizer.model_max_length:
            logger.warning(
                f"The block_size passed ({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(args.block_size, tokenizer.model_max_length)

    if not isinstance(grouping_strategy, str) and grouping_strategy == 1:
    # only keep the examples that are between 100 and 300 tokens
        validation_indices = [i for i, x in enumerate(tokenized_datasets["validation"]) if len(x['input_ids']) >= 100 and len(x['input_ids']) <= 300]
        test_indices = [i for i, x in enumerate(tokenized_datasets["test"]) if len(x['input_ids']) >= 100 and len(x['input_ids']) <= 300]
        tokenized_datasets["validation"] = tokenized_datasets["validation"].select(validation_indices)
        tokenized_datasets["test"] = tokenized_datasets["test"].select(test_indices)
    

    with accelerator.main_process_first():
        lm_datasets = tokenized_datasets.map(
            group_texts,
            batched=True,
            num_proc=args.preprocessing_num_workers,
            load_from_cache_file=not args.overwrite_cache,
            desc=f"Grouping texts in chunks of {block_size}",
            fn_kwargs={"tokenizer": tokenizer, "block_size": block_size,
                        "grouping_strategy": grouping_strategy, 
                        "padding_side": args.padding_side},
        )

    train_dataset = lm_datasets["train"]
    eval_dataset = lm_datasets["validation"]

    # Log a few random samples from the training set:
    for index in random.sample(range(len(train_dataset)), 3):
        logger.info(f"Sample {index} of the training set: {train_dataset[index]}.")

    # DataLoaders creation:
    train_dataloader = DataLoader(
        train_dataset, shuffle=True, collate_fn=default_data_collator, batch_size=args.per_device_train_batch_size
    )
    eval_dataloader = DataLoader(
        eval_dataset, collate_fn=default_data_collator, batch_size=args.per_device_eval_batch_size
    )

    # Optimizer
    freeze_param = []
    if args.freeze_emb == "gpt2_init":
        init_with_gpt2_emb(model, args)
    if args.freeze_emb is not None:
        freeze_param.append('transformer.wte.weight')
        freeze_param.append('transformer.wpe.weight')

    # Set the requires_grad=False for parameters that are to be frozen
    for n, p in model.named_parameters():
        if n in freeze_param:
            p.requires_grad = False
            print(n,p.shape,'frozen')
        else:
            print('trainable',n,p.shape)
    # filter out the non-trainable params based on requires_grad
    trainable_params = [(n, p) for n, p in model.named_parameters() if p.requires_grad]

    no_decay = ["bias", "layer_norm.weight"]
    optimizer_grouped_parameters = [
        {
            "params": [p for n, p in trainable_params if not any(nd in n for nd in no_decay)],
            "weight_decay": args.weight_decay,
        },
        {
            "params": [p for n, p in trainable_params if any(nd in n for nd in no_decay)],
            "weight_decay": 0.0,
        },
    ]
    optimizer = torch.optim.AdamW(optimizer_grouped_parameters, lr=args.learning_rate)

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

    if args.lr_scheduler_type != "lambdaLR":
        lr_scheduler = get_scheduler(
            name=args.lr_scheduler_type,
            optimizer=optimizer,
            num_warmup_steps=args.num_warmup_steps * accelerator.num_processes,
            num_training_steps=args.max_train_steps
            if overrode_max_train_steps
            else args.max_train_steps * accelerator.num_processes,
        )
    else:
        def lr_lambda(step):
            if step < args.dec_lr_epoch * num_update_steps_per_epoch:
                return 1.0
            else:
                return 0.1 # reduce the learning rate by 10x

        lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
            optimizer=optimizer, 
            lr_lambda=[lr_lambda, lr_lambda], 
            last_epoch=-1
        )

    # Prepare everything with our `accelerator`.
    model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
        model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
    )

    # On TPU, the tie weights in our model have been disconnected, so we need to restore the ties.
    if accelerator.distributed_type == DistributedType.TPU:
        model.tie_weights()

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # Figure out how many steps we should save the Accelerator states
    checkpointing_steps = args.checkpointing_steps
    epoch_save_steps = None
    if checkpointing_steps is not None and checkpointing_steps.isdigit():
        checkpointing_steps = int(checkpointing_steps)
    elif checkpointing_steps == "epoch":
        epoch_save_steps = get_epoch_save_steps(args.save_every_epoch, args.num_train_epochs)
    else:
        epoch_save_steps = None # don't save any epochs


    if args.with_tracking:
        wandb.login(key=WANDB_KEY)
        experiment_config = vars(args)
        # TensorBoard cannot log Enums, need the raw value
        # experiment_config["lr_scheduler_type"] = experiment_config["lr_scheduler_type"].value
        init_kwargs_logger = {}
        if args.wandb_run_name is not None:
            init_kwargs_logger = {"wandb": {"name": args.wandb_run_name, "entity": args.wandb_entity}}
        # TODO: change the name of the project in the following line for wandb
        accelerator.init_trackers(args.wandb_project_name, experiment_config, init_kwargs_logger)

    # Train!
    total_batch_size = args.per_device_train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.per_device_train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    logger.info(f"  Total update steps per epoch = {num_update_steps_per_epoch}")
    # Only show the progress bar once on each machine.
    progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
    completed_steps = 0
    starting_epoch = 0

    # Potentially load in the weights and states from a previous save
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint is not None or args.resume_from_checkpoint != "":
            checkpoint_path = args.resume_from_checkpoint
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = [f.name for f in os.scandir(os.getcwd()) if f.is_dir()]
            dirs.sort(key=os.path.getctime)
            path = dirs[-1]  # Sorts folders by date modified, most recent checkpoint is the last
            checkpoint_path = path
            path = os.path.basename(checkpoint_path)

        accelerator.print(f"Resumed from checkpoint: {checkpoint_path}")
        accelerator.load_state(checkpoint_path)
        # Extract `epoch_{i}` or `step_{i}`
        training_difference = os.path.splitext(path)[0]

        if "epoch" in training_difference:
            starting_epoch = int(training_difference.replace("epoch_", "")) + 1
            resume_step = None
            completed_steps = starting_epoch * num_update_steps_per_epoch
        else:
            # need to multiply `gradient_accumulation_steps` to reflect real steps
            resume_step = int(training_difference.replace("step_", "")) * args.gradient_accumulation_steps
            starting_epoch = resume_step // len(train_dataloader)
            completed_steps = resume_step // args.gradient_accumulation_steps
            resume_step -= starting_epoch * len(train_dataloader)

    # update the progress_bar if load from checkpoint
    progress_bar.update(completed_steps)

    interval_loss = 0
    interval_steps = 0
    prompt_seq2idx, gen_seq2idx = dump_seq2idx_mapping(train_dataset, tokenizer, args.output_dir)
    memories = {}
    unmem_seq_metrics = {}
    for epoch in range(starting_epoch, args.num_train_epochs):
        # print(f"epoch: {epoch}, lr: {lr_scheduler.get_last_lr()}")
        model.train()
        if args.with_tracking:
            total_loss = 0
        if args.resume_from_checkpoint and epoch == starting_epoch and resume_step is not None:
            # We skip the first `n` batches in the dataloader when resuming from a checkpoint
            active_dataloader = accelerator.skip_first_batches(train_dataloader, resume_step)
        else:
            active_dataloader = train_dataloader
        for step, batch in enumerate(active_dataloader):
            with accelerator.accumulate(model):
                outputs = model(**batch)
                loss = outputs.loss
                # We keep track of the loss at each epoch
                if args.with_tracking:
                    total_loss += loss.detach().float()
                    interval_loss += loss.detach().float()
                    interval_steps += 1
                accelerator.backward(loss)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                completed_steps += 1

            if isinstance(checkpointing_steps, int):
                if ((completed_steps % checkpointing_steps == 0) or (completed_steps == args.max_train_steps)) and accelerator.sync_gradients:
                    output_dir = f"step_{completed_steps}"
                    if args.output_dir is not None:
                        output_dir = os.path.join(args.output_dir, output_dir)
                    accelerator.save_state(output_dir)

            if ( ((completed_steps % args.eval_every_steps == 0) and (completed_steps != args.max_train_steps)) or completed_steps == 1):
                print_gen = False
                if completed_steps % args.print_gen_steps == 0:
                    print_gen = True
                    
                avg_interval_loss = interval_loss / interval_steps
                eval_loss, perplexity = do_eval_and_log(model, eval_dataloader, args, accelerator, epoch, completed_steps, logger, total_loss, avg_interval_loss, train_dataloader, train_dataset, tokenizer, eval_full_train=True, print_gen=print_gen, memories=memories, unmem_seq_metrics=unmem_seq_metrics, prompt_seq2idx=prompt_seq2idx, gen_seq2idx=gen_seq2idx)
                interval_loss = 0
                interval_steps = 0

            if (completed_steps == args.max_train_steps):
                avg_interval_loss = interval_loss / interval_steps
                eval_loss, perplexity = do_eval_and_log(model, eval_dataloader, args, accelerator, epoch, completed_steps, logger, total_loss, avg_interval_loss, train_dataloader, train_dataset, tokenizer, eval_full_train=True, print_gen=True, memories=memories, unmem_seq_metrics=unmem_seq_metrics, prompt_seq2idx=prompt_seq2idx, gen_seq2idx=gen_seq2idx)
                interval_loss = 0
                interval_steps = 0
                
            if completed_steps >= args.max_train_steps:
                os.makedirs(os.path.join(args.output_dir, "memories_metrics"), exist_ok=True)
                path = os.path.join(args.output_dir, "memories_metrics", f'memories.pkl')
                
                # add any summary metrics needed
                add_wandb_summary_metrics(accelerator, num_update_steps_per_epoch)
                
                # pickle dump the seq2idx mappings
                with open(path, 'wb') as f:
                    pickle.dump(memories, f)
                    
                path = os.path.join(args.output_dir, "memories_metrics", f'unmem_seq_metrics.pkl')
                with open(path, 'wb') as f:
                    pickle.dump(unmem_seq_metrics, f)
                
                break

        if args.checkpointing_steps == "epoch" and epoch_save_steps is not None:
            if ((epoch+1) % epoch_save_steps == 0) or ((epoch+1) == args.num_train_epochs):
                output_dir = f"epoch_{epoch+1}"
                if args.output_dir is not None:
                    output_dir = os.path.join(args.output_dir, output_dir)
                accelerator.save_state(output_dir)

    if args.with_tracking:
        accelerator.end_training()

    if args.output_dir is not None:
        accelerator.wait_for_everyone()
        unwrapped_model = accelerator.unwrap_model(model)
        unwrapped_model.save_pretrained(
            args.output_dir, is_main_process=accelerator.is_main_process, save_function=accelerator.save
        )
        if accelerator.is_main_process:
            tokenizer.save_pretrained(args.output_dir)
            if args.data_fraction >= 0.05:
                with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                    json.dump({"perplexity": perplexity}, f)
            else:
                with open(os.path.join(args.output_dir, "all_results.json"), "w") as f:
                    json.dump({"perplexity": math.exp(total_loss.item() / len(train_dataloader))}, f)


if __name__ == "__main__":
    main()