bash script for basic testing with pile-t5-large. note that this uses 1024 as the seq length for in/ 512 out

ft_flan.sh

#!/bin/bash

# Set environment variables
export WANDB_PROJECT="text2text-flan"
export WANDB_WATCH="gradients"
export WANDB_ENTITY="pszemraj"
export TOKENIZERS_PARALLELISM=true

NUM_WORKERS=$(lscpu -p | egrep -v '^#' | sort -u -t, -k 2,4 | wc -l)
echo "Number of CPU cores: $NUM_WORKERS"

# Set variables
MODEL_NAME_OR_PATH="pszemraj/tFINE-base-300m-SNI"
DATASET_NAME="pszemraj/flan-subsets-deduped"
DATASET_CONFIG="flan-v2"
TEXT_COLUMN="inputs"
SUMMARY_COLUMN="targets"
NUM_TRAIN_EPOCHS=1

LEARNING_RATE=8e-5
WARMUP_RATIO=0.05
OPTIM="adamw_torch_fused" # paged_adamw_32bit
LR_SCHEDULER_TYPE="inverse_sqrt"
WEIGHT_DECAY=0.04
MAX_GRAD_NORM=1.0

PER_DEVICE_TRAIN_BATCH_SIZE=4
PER_DEVICE_EVAL_BATCH_SIZE=4
GRADIENT_ACCUMULATION_STEPS=32

MAX_SOURCE_LENGTH=1024
MAX_TARGET_LENGTH=1024
VAL_MAX_TARGET_LENGTH=1024
GENERATION_MAX_LENGTH=1024
NUM_BEAMS=1

SEED=17868
DATA_SEED=16919
MODEL_NAME=$(basename "$MODEL_NAME_OR_PATH")
DS_BASENAME=$(basename "$DATASET_NAME")
RUN_NAME="$MODEL_NAME-$DS_BASENAME-$MAX_SOURCE_LENGTH"
HUB_MODEL_ID="pszemraj/$RUN_NAME"
OUTPUT_DIR="./outputs/$RUN_NAME"

EVAL_STRATEGY="steps"
EVAL_STEPS=1000
MAX_EVAL_SAMPLES=150
LOGGING_DIR="./$OUTPUT_DIR/logs"
LOGGING_STEPS=10
SAVE_STRATEGY="steps"

accelerate launch ./run_summarization.py \
  --model_name_or_path "$MODEL_NAME_OR_PATH" \
  --do_train \
  --do_eval \
  --evaluation_strategy "$EVAL_STRATEGY" --eval_steps "$EVAL_STEPS" \
  --dataset_name "$DATASET_NAME" --dataset_config_name "$DATASET_CONFIG" \
  --text_column "$TEXT_COLUMN" \
  --summary_column "$SUMMARY_COLUMN" \
  --bf16 \
  --bf16_full_eval False \
  --dataloader_num_workers "$NUM_WORKERS" \
  --filter True \
  --generation_max_length "$GENERATION_MAX_LENGTH" \
  --gradient_accumulation_steps "$GRADIENT_ACCUMULATION_STEPS" \
  --gradient_checkpointing True \
  --hub_model_id "$HUB_MODEL_ID" \
  --hub_private_repo True \
  --hub_strategy every_save \
  --include_num_input_tokens_seen True \
  --learning_rate "$LEARNING_RATE" \
  --logging_dir "$LOGGING_DIR" \
  --logging_steps "$LOGGING_STEPS" \
  --lr_scheduler_type "$LR_SCHEDULER_TYPE" \
  --max_eval_samples "$MAX_EVAL_SAMPLES" \
  --max_grad_norm "$MAX_GRAD_NORM" \
  --max_source_length "$MAX_SOURCE_LENGTH" \
  --max_target_length "$MAX_TARGET_LENGTH" \
  --num_beams "$NUM_BEAMS" \
  --num_train_epochs "$NUM_TRAIN_EPOCHS" \
  --optim "$OPTIM" \
  --output_dir "$OUTPUT_DIR" \
  --overwrite_output_dir True \
  --pad_to_max_length False \
  --per_device_eval_batch_size "$PER_DEVICE_EVAL_BATCH_SIZE" \
  --per_device_train_batch_size "$PER_DEVICE_TRAIN_BATCH_SIZE" \
  --predict_with_generate True \
  --preprocessing_num_workers "$NUM_WORKERS" \
  --push_to_hub \
  --report_to wandb \
  --run_name "$RUN_NAME" \
  --save_total_limit 1 \
  --seed "$SEED" \
  --sortish_sampler True \
  --tf32 True \
  --val_max_target_length "$VAL_MAX_TARGET_LENGTH" \
  --warmup_ratio "$WARMUP_RATIO" \
  --weight_decay "$WEIGHT_DECAY" \
  --torch_compile_backend "inductor"


# --use_fast_tokenizer False

map_reduce.sh

#!/bin/bash

# Set environment variables
export WANDB_PROJECT="summary-map-reduce"
export WANDB_WATCH="gradients"
export WANDB_ENTITY="pszemraj"
export TOKENIZERS_PARALLELISM=true

NUM_WORKERS=$(lscpu -p | egrep -v '^#' | sort -u -t, -k 2,4 | wc -l)
echo "Number of CPU cores: $NUM_WORKERS"

# Set variables
MODEL_NAME_OR_PATH="google/flan-t5-large"
DATASET_NAME="pszemraj/summary-map-reduce"
DATASET_CONFIG="all-filtered"
TEXT_COLUMN="batched_summary"
SUMMARY_COLUMN="combined_summary"
NUM_TRAIN_EPOCHS=2

LEARNING_RATE=1e-4
WARMUP_RATIO=0.05
OPTIM="paged_adamw_8bit" #"paged_adamw_32bit"
LR_SCHEDULER_TYPE="cosine"
WEIGHT_DECAY=0.01
MAX_GRAD_NORM=1.0

PER_DEVICE_TRAIN_BATCH_SIZE=4
PER_DEVICE_EVAL_BATCH_SIZE=4
GRADIENT_ACCUMULATION_STEPS=16

MAX_SOURCE_LENGTH=1024
MAX_TARGET_LENGTH=1024
VAL_MAX_TARGET_LENGTH=1024
GENERATION_MAX_LENGTH=1024
NUM_BEAMS=1

SEED=17868
DATA_SEED=16919
MODEL_NAME=$(basename "$MODEL_NAME_OR_PATH")
DS_BASENAME=$(basename "$DATASET_NAME")
RUN_NAME="$MODEL_NAME-$DS_BASENAME-$MAX_SOURCE_LENGTH"
HUB_MODEL_ID="pszemraj/$RUN_NAME"
OUTPUT_DIR="./outputs/$RUN_NAME"

EVAL_STRATEGY="steps"
EVAL_STEPS=100
MAX_EVAL_SAMPLES=150
LOGGING_DIR="./$OUTPUT_DIR/logs"
LOGGING_STEPS=5
SAVE_STRATEGY="steps"

python ./run_summarization.py \
  --model_name_or_path "$MODEL_NAME_OR_PATH" \
  --do_train \
  --do_eval \
  --evaluation_strategy "$EVAL_STRATEGY" --eval_steps "$EVAL_STEPS" \
  --dataset_name "$DATASET_NAME" --dataset_config_name "$DATASET_CONFIG" \
  --text_column "$TEXT_COLUMN" \
  --summary_column "$SUMMARY_COLUMN" \
  --bf16 \
  --bf16_full_eval False \
  --dataloader_num_workers "$NUM_WORKERS" \
  --filter True \
  --generation_max_length "$GENERATION_MAX_LENGTH" \
  --gradient_accumulation_steps "$GRADIENT_ACCUMULATION_STEPS" \
  --gradient_checkpointing True \
  --hub_model_id "$HUB_MODEL_ID" \
  --hub_private_repo True \
  --hub_strategy every_save \
  --include_num_input_tokens_seen True \
  --learning_rate "$LEARNING_RATE" \
  --logging_dir "$LOGGING_DIR" \
  --logging_steps "$LOGGING_STEPS" \
  --lr_scheduler_type "$LR_SCHEDULER_TYPE" \
  --max_eval_samples "$MAX_EVAL_SAMPLES" \
  --max_grad_norm "$MAX_GRAD_NORM" \
  --max_source_length "$MAX_SOURCE_LENGTH" \
  --max_target_length "$MAX_TARGET_LENGTH" \
  --num_beams "$NUM_BEAMS" \
  --num_train_epochs "$NUM_TRAIN_EPOCHS" \
  --optim "$OPTIM" \
  --output_dir "$OUTPUT_DIR" \
  --overwrite_output_dir True \
  --pad_to_max_length False \
  --per_device_eval_batch_size "$PER_DEVICE_EVAL_BATCH_SIZE" \
  --per_device_train_batch_size "$PER_DEVICE_TRAIN_BATCH_SIZE" \
  --predict_with_generate False \
  --preprocessing_num_workers "$NUM_WORKERS" \
  --push_to_hub \
  --report_to wandb \
  --run_name "$RUN_NAME" \
  --save_total_limit 1 --save_steps 100 \
  --seed "$SEED" \
  --sortish_sampler True \
  --tf32 True \
  --val_max_target_length "$VAL_MAX_TARGET_LENGTH" \
  --warmup_ratio "$WARMUP_RATIO" \
  --weight_decay "$WEIGHT_DECAY" \
  --torch_compile_backend "inductor"


# --use_fast_tokenizer False

pile_t5_samsum.sh

#!/bin/bash

# Set environment variables
export WANDB_PROJECT="pileT5-summ"
export WANDB_WATCH="gradients"
export WANDB_ENTITY="pszemraj"

NUM_WORKERS=$(lscpu -p | egrep -v '^#' | sort -u -t, -k 2,4 | wc -l)
echo "Number of CPU cores: $NUM_WORKERS"

# Set variables
MODEL_NAME_OR_PATH="pszemraj/nanoT5-base-SiLU"
DATASET_NAME="samsum"
TEXT_COLUMN="dialogue"
SUMMARY_COLUMN="summary"
NUM_TRAIN_EPOCHS=5

LEARNING_RATE=1e-4
WARMUP_RATIO=0.05
OPTIM="adamw_torch_fused"
LR_SCHEDULER_TYPE="cosine"
WEIGHT_DECAY=0.01
MAX_GRAD_NORM=1.0

PER_DEVICE_TRAIN_BATCH_SIZE=8
PER_DEVICE_EVAL_BATCH_SIZE=16
GRADIENT_ACCUMULATION_STEPS=16

MAX_SOURCE_LENGTH=1024
MAX_TARGET_LENGTH=512
VAL_MAX_TARGET_LENGTH=512
VAL_MAX_TARGET_LENGTH=512
GENERATION_MAX_LENGTH=512
NUM_BEAMS=1

SEED=17868
DATA_SEED=16919
MODEL_NAME=$(basename "$MODEL_NAME_OR_PATH")
DS_BASENAME=$(basename "$DATASET_NAME")
RUN_NAME="$MODEL_NAME-$DS_BASENAME"
HUB_MODEL_ID="BEE-spoke-data/$RUN_NAME"
OUTPUT_DIR="./runtime/$RUN_NAME"
LOGGING_DIR="./runtime/$RUN_NAME/logs"
LOGGING_STEPS=5
MAX_EVAL_SAMPLES=300
METRIC_FOR_BEST_MODEL="rouge2"
SAVE_STRATEGY="epoch"

python ./run_summarization.py \
  --model_name_or_path "$MODEL_NAME_OR_PATH" \
  --do_train \
  --do_eval \
  --do_predict \
  --evaluation_strategy epoch \
  --dataset_name "$DATASET_NAME" \
  --text_column "$TEXT_COLUMN" \
  --summary_column "$SUMMARY_COLUMN" \
  --bf16 \
  --bf16_full_eval False \
  --data_seed "$DATA_SEED" \
  --dataloader_num_workers "$NUM_WORKERS" \
  --preprocessing_num_workers "$NUM_WORKERS" \
  --generation_max_length "$GENERATION_MAX_LENGTH" \
  --gradient_accumulation_steps "$GRADIENT_ACCUMULATION_STEPS" \
  --gradient_checkpointing True \
  --learning_rate "$LEARNING_RATE" \
  --load_best_model_at_end True \
  --logging_dir "$LOGGING_DIR" \
  --logging_steps "$LOGGING_STEPS" \
  --lr_scheduler_type "$LR_SCHEDULER_TYPE" \
  --max_eval_samples "$MAX_EVAL_SAMPLES" \
  --max_grad_norm "$MAX_GRAD_NORM" \
  --max_source_length "$MAX_SOURCE_LENGTH" \
  --max_target_length "$MAX_TARGET_LENGTH" \
  --metric_for_best_model "$METRIC_FOR_BEST_MODEL" \
  --num_beams "$NUM_BEAMS" \
  --num_train_epochs "$NUM_TRAIN_EPOCHS" \
  --optim "$OPTIM" \
  --output_dir "$OUTPUT_DIR" \
  --overwrite_output_dir True \
  --pad_to_max_length False \
  --per_device_eval_batch_size "$PER_DEVICE_EVAL_BATCH_SIZE" \
  --per_device_train_batch_size "$PER_DEVICE_TRAIN_BATCH_SIZE" \
  --predict_with_generate True \
  --report_to wandb \
  --run_name "$RUN_NAME" \
  --save_strategy "$SAVE_STRATEGY" \
  --seed "$SEED" \
  --sortish_sampler True \
  --tf32 True \
  --val_max_target_length "$VAL_MAX_TARGET_LENGTH" \
  --warmup_ratio "$WARMUP_RATIO" \
  --weight_decay "$WEIGHT_DECAY" \
  --greater_is_better True \
  --torch_compile_backend "inductor"

# --hub_model_id "$HUB_MODEL_ID" \
  # --hub_private_repo True \
  # --hub_strategy every_save \
  # --push_to_hub \
#   --use_fast_tokenizer False

run_summarization.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2021 The HuggingFace 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 sequence to sequence.
"""

import inspect
import json
import logging
import os
import sys
import time
import warnings
from dataclasses import dataclass, field
from typing import Optional
from packaging.version import Version

import datasets
import evaluate
import nltk  # Here to have a nice missing dependency error message early on
import numpy as np
import pandas as pd
import torch
import transformers
from datasets import Dataset, load_dataset
from filelock import FileLock
from torch.nn.attention import SDPBackend, sdpa_kernel
from transformers import (
    AutoConfig,
    AutoModelForSeq2SeqLM,
    AutoTokenizer,
    DataCollatorForSeq2Seq,
    HfArgumentParser,
    MBart50Tokenizer,
    MBart50TokenizerFast,
    MBartTokenizer,
    MBartTokenizerFast,
    Seq2SeqTrainer,
    Seq2SeqTrainingArguments,
    set_seed,
)
from transformers.trainer_utils import get_last_checkpoint
from transformers.utils import is_offline_mode
from transformers.utils.versions import require_version

# =============================================
# torch.cuda.amp.custom_fwd is deprecated >= 2.4
torch_version = torch.__version__
if Version(torch_version) < Version("2.4.0"):
    torch_amp_custom_fwd = torch.cuda.amp.custom_fwd
    torch_amp_custom_bwd = torch.cuda.amp.custom_bwd
else:
    torch_amp_custom_fwd = torch.amp.custom_fwd(device_type = "cuda")
    torch_amp_custom_bwd = torch.amp.custom_bwd(device_type = "cuda")
pass
# =============================================

# Filter out specific warnings about tensor resizing
warnings.filterwarnings(
    "ignore",
    message=".*An output with one or more elements was resized since it had shape.*",
    category=UserWarning,
    module="transformers.models.longt5.modeling_longt5",
)


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

logger = logging.getLogger(__name__)

try:
    nltk.data.find("tokenizers/punkt")
except (LookupError, OSError):
    if is_offline_mode():
        raise LookupError(
            "Offline mode: run this script without TRANSFORMERS_OFFLINE first to download nltk data files"
        )
    with FileLock(".lock") as lock:
        nltk.download("punkt_tab", quiet=True)
        nltk.download("popular", quiet=True) # they keep renaming their files

# A list of all multilingual tokenizer which require lang attribute.
MULTILINGUAL_TOKENIZERS = [
    MBartTokenizer,
    MBartTokenizerFast,
    MBart50Tokenizer,
    MBart50TokenizerFast,
]

# >>> DYNAMO UPDATES

# Torch compile arguments
torch_compile_arguments = [
    "config.dce = True",
    "config.memory_planning = True",
    "config.memory_pool = 'combined'",
    "config.coordinate_descent_tuning = True",
    "config.max_autotune_gemm = False", # GEMM is unnecessary
    "config.autotune_multi_device = False",
    "config.max_autotune_gemm_backends = 'ATEN'", # Not much faster
    "config.aggressive_fusion = False", # Careful changes results!
    "config.cuda.enable_cuda_lto = True",
    "config.cuda.use_fast_math = True",
    "config.cuda.compile_opt_level = '-O3'",
]
# Torch dynamo arguments
torch_dynamo_arguments = [
    "config.accumulated_cache_size_limit = 512", # Bump up a bit from 256
    "config.suppress_errors = True", # Supress errors for now
    "config.do_not_emit_runtime_asserts = True",
]
import torch._inductor.config as config
for _try_compile_argument in torch_compile_arguments:
    try:    exec(_try_compile_argument)
    except: pass
pass
import torch._dynamo.config as config
for _try_dynamo_argument in torch_dynamo_arguments:
    try:    exec(_try_dynamo_argument)
    except: pass
pass

# >>> DYNAMO UPDATES

class Unsloth_Offloaded_Gradient_Checkpointer(torch.autograd.Function):
    """
    Saves VRAM by smartly offloading to RAM.

    - Tiny hit to performance, since we mask the movement via non blocking calls.
    - adapted from unsloth fn for encoder-decoder models
    """
    @staticmethod
    @torch.cuda.amp.custom_fwd
    def forward(ctx, forward_function, *args):
        ctx.forward_function = forward_function
        ctx.non_tensor_indices = []
        saved_tensors = []

        for i, arg in enumerate(args):
            if isinstance(arg, torch.Tensor):
                saved_tensors.append(arg.to("cpu", non_blocking=True))
            else:
                ctx.non_tensor_indices.append(i)
                saved_tensors.append(arg)

        ctx.save_for_backward(*saved_tensors)

        with torch.no_grad():
            output = forward_function(*args)
        return output

    @staticmethod
    @torch.cuda.amp.custom_bwd
    def backward(ctx, *grad_outputs):
        args = list(ctx.saved_tensors)
        for i, arg in enumerate(args):
            if i not in ctx.non_tensor_indices:
                args[i] = arg.to("cuda:0", non_blocking=True).detach()
                args[i].requires_grad = True

        with torch.enable_grad():
            output = ctx.forward_function(*args)

        if not isinstance(output, tuple):
            output = (output,)

        torch.autograd.backward(output, grad_outputs)

        grads = []
        for i, arg in enumerate(args):
            if i in ctx.non_tensor_indices:
                grads.append(None)
            else:
                grads.append(arg.grad)

        return (None,) + tuple(grads)

@torch._disable_dynamo
def unsloth_offloaded_gradient_checkpoint(function, *args, use_reentrant=None, **kwargs):
    return Unsloth_Offloaded_Gradient_Checkpointer.apply(function, *args)

from transformers import PreTrainedModel

def apply_unsloth_offloaded_gradient_checkpoint_monkey_patch():
    """
    Monkey patches the Transformers library to use Unsloth's offloaded gradient checkpointing
    for encoder-decoder models, allowing additional keyword arguments.
    """

    print("Applying offloaded gradient checkpointing monkey patch...")
    # Check if the model supports gradient checkpointing
    def supports_gradient_checkpointing(module):
        return hasattr(module, "gradient_checkpointing_enable")

    # New function for setting up offloaded gradient checkpointing
    def unsloth_gradient_checkpointing(module, use_reentrant=True, **kwargs):
        if not supports_gradient_checkpointing(module):
            raise TypeError("The provided model does not support gradient checkpointing.")

        # Enable gradient checkpointing using the custom offloaded function
        def patched_forward(*inputs, **kwargs):
            def custom_forward(*inputs):
                return module.forward(*inputs, **kwargs)

            # Use the custom offloaded gradient checkpointing function
            return unsloth_offloaded_gradient_checkpoint(custom_forward, *inputs, **kwargs)

        module.forward = patched_forward

    # Patch the encoder and decoder models
    def patch_encoder_decoder(model, **kwargs):
        if hasattr(model, "encoder"):
            unsloth_gradient_checkpointing(model.encoder, **kwargs)

        if hasattr(model, "decoder"):
            unsloth_gradient_checkpointing(model.decoder, **kwargs)

    # Patch the model's gradient checkpointing functionality
    def patch_model(model: PreTrainedModel, **kwargs):
        if supports_gradient_checkpointing(model):
            original_enable = getattr(model, 'gradient_checkpointing_enable', None)
            def new_enable(**enable_kwargs):
                if original_enable is not None:
                    original_enable()  # Call the original enable function
                patch_encoder_decoder(model, **{**kwargs, **enable_kwargs})
            model.gradient_checkpointing_enable = new_enable
        else:
            raise TypeError("The provided model does not support gradient checkpointing.")
    # Apply the patch globally to all models
    original_init = PreTrainedModel.__init__

    def patched_init(self, *args, **kwargs):
        original_init(self, *args, **kwargs)
        if supports_gradient_checkpointing(self):
            self.gradient_checkpointing_enable = lambda **kwargs: patch_model(self, **kwargs)

    PreTrainedModel.__init__ = patched_init

# Apply the monkey patch
# apply_unsloth_offloaded_gradient_checkpoint_monkey_patch()

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

    model_name_or_path: str = field(
        metadata={
            "help": "Path to pretrained model or model identifier from huggingface.co/models"
        }
    )
    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 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)."
        },
    )
    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)."
            )
        },
    )
    resize_position_embeddings: Optional[bool] = field(
        default=None,
        metadata={
            "help": (
                "Whether to automatically resize the position embeddings if `max_source_length` exceeds "
                "the model's position embeddings."
            )
        },
    )
    disable_sdp_attention: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Disable the Scaled Dot-Product Attention (SDP) torch context manager."
        },
    )
    update_activation_fn: Optional[bool] = field(
        default=False,
        metadata={
            "help": "Update the model's activation function to silu if not already set."
        },
    )


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

    lang: Optional[str] = field(
        default=None, metadata={"help": "Language id for summarization."}
    )

    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)."
        },
    )
    text_column: Optional[str] = field(
        default=None,
        metadata={
            "help": "The name of the column in the datasets containing the full texts (for summarization)."
        },
    )
    summary_column: Optional[str] = field(
        default=None,
        metadata={
            "help": "The name of the column in the datasets containing the summaries (for summarization)."
        },
    )
    train_file: Optional[str] = field(
        default=None,
        metadata={"help": "The input training data file (a jsonlines or csv file)."},
    )
    validation_file: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "An optional input evaluation data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
            )
        },
    )
    test_file: Optional[str] = field(
        default=None,
        metadata={
            "help": "An optional input test data file to evaluate the metrics (rouge) on (a jsonlines or csv file)."
        },
    )
    overwrite_cache: bool = field(
        default=False,
        metadata={"help": "Overwrite the cached training and evaluation sets"},
    )
    preprocessing_num_workers: Optional[int] = field(
        default=None,
        metadata={"help": "The number of processes to use for the preprocessing."},
    )
    max_source_length: Optional[int] = field(
        default=1024,
        metadata={
            "help": (
                "The maximum total input sequence length after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    max_target_length: Optional[int] = field(
        default=128,
        metadata={
            "help": (
                "The maximum total sequence length for target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded."
            )
        },
    )
    val_max_target_length: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "The maximum total sequence length for validation target text after tokenization. Sequences longer "
                "than this will be truncated, sequences shorter will be padded. Will default to `max_target_length`."
                "This argument is also used to override the ``max_length`` param of ``model.generate``, which is used "
                "during ``evaluate`` and ``predict``."
            )
        },
    )
    pad_to_max_length: bool = field(
        default=False,
        metadata={
            "help": (
                "Whether to pad all samples to model maximum sentence length. "
                "If False, will pad the samples dynamically when batching to the maximum length in the batch. More "
                "efficient on GPU but very bad for TPU."
            )
        },
    )
    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."
            )
        },
    )
    max_predict_samples: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "For debugging purposes or quicker training, truncate the number of prediction examples to this "
                "value if set."
            )
        },
    )
    num_beams: Optional[int] = field(
        default=None,
        metadata={
            "help": (
                "Number of beams to use for evaluation. This argument will be passed to ``model.generate``, "
                "which is used during ``evaluate`` and ``predict``."
            )
        },
    )
    ignore_pad_token_for_loss: bool = field(
        default=True,
        metadata={
            "help": "Whether to ignore the tokens corresponding to padded labels in the loss computation or not."
        },
    )
    source_prefix: Optional[str] = field(
        default="",
        metadata={
            "help": "A prefix to add before every source text (useful for T5 models)."
        },
    )

    forced_bos_token: Optional[str] = field(
        default=None,
        metadata={
            "help": (
                "The token to force as the first generated token after the decoder_start_token_id."
                "Useful for multilingual models like mBART where the first generated token"
                "needs to be the target language token (Usually it is the target language token)"
            )
        },
    )
    shuffle: bool = field(
        default=False,
        metadata={
            "help": "Shuffle the dataset right before training using the run seed."
        },
    )
    do_predict_testset: bool = field(
        default=False,
        metadata={
            "help": "Whether to run prediction on the test set. (as originally done by this script)"
        },
    )
    filter: bool = field(
        default=False,
        metadata={
            "help": "Whether to filter the dataset based on input/output lengths using max_source_length and max_target_length"
        },
    )

    def __post_init__(self):
        if (
            self.dataset_name is None
            and self.train_file is None
            and self.validation_file is None
            and self.test_file is None
        ):
            raise ValueError(
                "Need either a dataset name or a training, validation, or test file."
            )
        else:
            if self.train_file is not None:
                extension = self.train_file.split(".")[-1]
                assert extension in [
                    "csv",
                    "json",
                ], "`train_file` should be a csv or a json file."
            if self.validation_file is not None:
                extension = self.validation_file.split(".")[-1]
                assert extension in [
                    "csv",
                    "json",
                ], "`validation_file` should be a csv or a json file."
            if self.test_file is not None:
                extension = self.test_file.split(".")[-1]
                assert extension in [
                    "csv",
                    "json",
                ], "`test_file` should be a csv or a json file."
        if self.val_max_target_length is None:
            self.val_max_target_length = self.max_target_length


summarization_name_mapping = {
    "amazon_reviews_multi": ("review_body", "review_title"),
    "big_patent": ("description", "abstract"),
    "cnn_dailymail": ("article", "highlights"),
    "orange_sum": ("text", "summary"),
    "pn_summary": ("article", "summary"),
    "psc": ("extract_text", "summary_text"),
    "samsum": ("dialogue", "summary"),
    "thaisum": ("body", "summary"),
    "xglue": ("news_body", "news_title"),
    "xsum": ("document", "summary"),
    "wiki_summary": ("article", "highlights"),
    "multi_news": ("document", "summary"),
}


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, Seq2SeqTrainingArguments)
    )
    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()

    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:
        transformers.utils.logging.set_verbosity_info()

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

    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}")

    if data_args.source_prefix is None and model_args.model_name_or_path in [
        "t5-small",
        "t5-base",
        "t5-large",
        "t5-3b",
        "t5-11b",
    ]:
        logger.warning(
            "You're running a t5 model but didn't provide a source prefix, which is the expected, e.g. with "
            "`--source_prefix 'summarize: ' `"
        )

    # 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 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 first column for the full texts and the second column for the
    # summaries (unless you specify column names for this with the `text_column` and `summary_column` arguments).
    #
    # 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,
            num_proc=data_args.preprocessing_num_workers,
            trust_remote_code=True,
        )
    else:
        data_files = {}
        if data_args.train_file is not None:
            data_files["train"] = data_args.train_file
            extension = data_args.train_file.split(".")[-1]
        if data_args.validation_file is not None:
            data_files["validation"] = data_args.validation_file
            extension = data_args.validation_file.split(".")[-1]
        if data_args.test_file is not None:
            data_files["test"] = data_args.test_file
            extension = data_args.test_file.split(".")[-1]
        raw_datasets = load_dataset(
            extension,
            data_files=data_files,
            cache_dir=model_args.cache_dir,
            use_auth_token=True if model_args.use_auth_token else None,
        )

    if "validation" not in raw_datasets:
        logger.warning("No validation split found, creating one from the training data")
        train_val_split = raw_datasets["train"].train_test_split(
            test_size=0.05, seed=training_args.seed  # 5% for validation
        )
        raw_datasets["train"] = train_val_split["train"]
        raw_datasets["validation"] = train_val_split["test"]

    # 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 = AutoConfig.from_pretrained(
        (
            model_args.config_name
            if model_args.config_name
            else model_args.model_name_or_path
        ),
        cache_dir=model_args.cache_dir,
        revision=model_args.model_revision,
        use_auth_token=True if model_args.use_auth_token else None,
    )
    tokenizer = AutoTokenizer.from_pretrained(
        (
            model_args.tokenizer_name
            if model_args.tokenizer_name
            else model_args.model_name_or_path
        ),
        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,
    )
    if model_args.update_activation_fn and config.activation_function != "silu":
        logger.info(
            f"Update activation function from {config.activation_function} to silu"
        )
        config.activation_function = "silu"

    model = AutoModelForSeq2SeqLM.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,
    )
    logger.info(f"Model {model_args.model_name_or_path} loaded in dtype {model.dtype}")
    logger.info(f"Input model config: {model.config}")
    time.sleep(2)

    # 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 or (embedding_size % 128) != 0:
        logger.info(
            f"Resizing model embedding size from {embedding_size} to "
            f"{len(tokenizer)} & pad_to_multiple_of=128"
        )
        model.resize_token_embeddings(len(tokenizer), pad_to_multiple_of=128)
        logger.info(f"New vocab size is {model.get_input_embeddings().weight.shape[0]}")
        model.config.vocab_size = model.get_input_embeddings().weight.shape[0]

    if model.config.decoder_start_token_id is None and isinstance(
        tokenizer, (MBartTokenizer, MBartTokenizerFast)
    ):
        if isinstance(tokenizer, MBartTokenizer):
            model.config.decoder_start_token_id = tokenizer.lang_code_to_id[
                data_args.lang
            ]
        else:
            model.config.decoder_start_token_id = tokenizer.convert_tokens_to_ids(
                data_args.lang
            )

    if model.config.decoder_start_token_id is None:
        raise ValueError(
            "Make sure that `config.decoder_start_token_id` is correctly defined"
        )

    if (
        hasattr(model.config, "max_position_embeddings")
        and model.config.max_position_embeddings < data_args.max_source_length
    ):
        if model_args.resize_position_embeddings is None:
            logger.warning(
                "Increasing the model's number of position embedding vectors from"
                f" {model.config.max_position_embeddings} to {data_args.max_source_length}."
            )
            model.resize_position_embeddings(data_args.max_source_length)
        elif model_args.resize_position_embeddings:
            model.resize_position_embeddings(data_args.max_source_length)
        else:
            raise ValueError(
                f"`--max_source_length` is set to {data_args.max_source_length}, but the model only has"
                f" {model.config.max_position_embeddings} position encodings. Consider either reducing"
                f" `--max_source_length` to {model.config.max_position_embeddings} or to automatically resize the"
                " model's position encodings by passing `--resize_position_embeddings`."
            )

    prefix = data_args.source_prefix if data_args.source_prefix is not None else ""

    # Preprocessing the datasets.
    if training_args.do_train:
        if "train" not in raw_datasets:
            raise ValueError("--do_train requires a train dataset")
        column_names = raw_datasets["train"].column_names
    elif training_args.do_eval:
        if "validation" not in raw_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        column_names = raw_datasets["validation"].column_names
    elif training_args.do_predict:
        if "test" not in raw_datasets:
            raise ValueError("--do_predict requires a test dataset")
        column_names = raw_datasets["test"].column_names
    else:
        logger.info(
            "There is nothing to do. Please pass `do_train`, `do_eval` and/or `do_predict`."
        )
        return

    if isinstance(tokenizer, tuple(MULTILINGUAL_TOKENIZERS)):
        assert (
            data_args.lang is not None
        ), f"{tokenizer.__class__.__name__} is a multilingual tokenizer which requires --lang argument"

        tokenizer.src_lang = data_args.lang
        tokenizer.tgt_lang = data_args.lang

        # For multilingual translation models like mBART-50 and M2M100 we need to force the target language token
        # as the first generated token. We ask the user to explicitly provide this as --forced_bos_token argument.
        forced_bos_token_id = (
            tokenizer.lang_code_to_id[data_args.forced_bos_token]
            if data_args.forced_bos_token is not None
            else None
        )
        model.config.forced_bos_token_id = forced_bos_token_id

    # Get the column names for input/target.
    dataset_columns = summarization_name_mapping.get(data_args.dataset_name, None)
    if data_args.text_column is None:
        text_column = (
            dataset_columns[0] if dataset_columns is not None else column_names[0]
        )
    else:
        text_column = data_args.text_column
        if text_column not in column_names:
            raise ValueError(
                f"--text_column' value '{data_args.text_column}' needs to be one of: {', '.join(column_names)}"
            )
    if data_args.summary_column is None:
        summary_column = (
            dataset_columns[1] if dataset_columns is not None else column_names[1]
        )
    else:
        summary_column = data_args.summary_column
        if summary_column not in column_names:
            raise ValueError(
                f"--summary_column' value '{data_args.summary_column}' needs to be one of: {', '.join(column_names)}"
            )

    # Temporarily set max_target_length for training.
    max_target_length = data_args.max_target_length
    padding = "max_length" if data_args.pad_to_max_length else False

    if training_args.label_smoothing_factor > 0 and not hasattr(
        model, "prepare_decoder_input_ids_from_labels"
    ):
        logger.warning(
            "label_smoothing is enabled but the `prepare_decoder_input_ids_from_labels` method is not defined for"
            f"`{model.__class__.__name__}`. This will lead to loss being calculated twice and will take up more memory"
        )

    def filter_dataset(
        dataset,
        tokenizer,
        text_column,
        summary_column,
        max_input_length,
        max_output_length,
    ) -> Dataset:
        """
        filter_dataset - filter dataset based on max_input_length and max_output_length
        """

        def check_length(example):
            input_length = len(
                tokenizer.encode(example[text_column], padding=False, truncation=False)
            )
            output_length = len(
                tokenizer.encode(
                    example[summary_column], padding=False, truncation=False
                )
            )

            keep = True
            if max_input_length and input_length > max_input_length:
                keep = False
            if max_output_length and output_length > max_output_length:
                keep = False

            return keep

        original_num_examples = len(dataset)
        filtered_dataset = dataset.filter(
            check_length,
            num_proc=data_args.preprocessing_num_workers,
            desc="filtering long examples",
        )
        filtered_num_examples = len(filtered_dataset)

        logger.info(
            f"Filtered: keep {filtered_num_examples} / {original_num_examples} original"
        )

        return filtered_dataset

    def preprocess_function(examples):
        # remove pairs where at least one record is None

        inputs, targets = [], []
        for i in range(len(examples[text_column])):
            if examples[text_column][i] and examples[summary_column][i]:
                inputs.append(examples[text_column][i])
                targets.append(examples[summary_column][i])

        inputs = [prefix + inp for inp in inputs]
        model_inputs = tokenizer(
            inputs,
            max_length=data_args.max_source_length,
            padding=padding,
            truncation=True,
        )

        # Tokenize targets with the `text_target` keyword argument
        labels = tokenizer(
            text_target=targets,
            max_length=max_target_length,
            padding=padding,
            truncation=True,
        )

        # If we are padding here, replace all tokenizer.pad_token_id in the labels by -100
        # when we want to ignore padding in the loss.
        if padding == "max_length" and data_args.ignore_pad_token_for_loss:
            labels["input_ids"] = [
                [(l if l != tokenizer.pad_token_id else -100) for l in label]
                for label in labels["input_ids"]
            ]

        model_inputs["labels"] = labels["input_ids"]
        return model_inputs

    if training_args.do_train:
        if "train" not in raw_datasets:
            raise ValueError("--do_train requires a train dataset")
        train_dataset = raw_datasets["train"]

        if data_args.filter:
            logger.info("Filtering train dataset based on input/output lengths")
            train_dataset = filter_dataset(
                train_dataset,
                tokenizer,
                text_column,
                summary_column,
                data_args.max_source_length,
                data_args.max_target_length,
            )

        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))

        with training_args.main_process_first(desc="train dataset map pre-processing"):
            train_dataset = train_dataset.map(
                preprocess_function,
                batched=True,
                batch_size=128 if data_args.preprocessing_num_workers > 30 else 512,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on train dataset",
            )

        logger.info(f"shuffling train dataset with seed {training_args.seed}")
        train_dataset = train_dataset.shuffle(training_args.seed)

    if training_args.do_eval:
        if "validation" not in raw_datasets:
            raise ValueError("--do_eval requires a validation dataset")
        eval_dataset = raw_datasets["validation"]

        if data_args.filter:
            logger.info("Filtering validation dataset based on input/output lengths")
            eval_dataset = filter_dataset(
                eval_dataset,
                tokenizer,
                text_column,
                summary_column,
                data_args.max_source_length,
                data_args.max_target_length,
            )

        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))

        with training_args.main_process_first(
            desc="validation dataset map pre-processing"
        ):
            eval_dataset = eval_dataset.map(
                preprocess_function,
                batched=True,
                batch_size=128 if data_args.preprocessing_num_workers > 30 else 512,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on validation dataset",
            )
    if training_args.do_predict:
        max_target_length = data_args.val_max_target_length
        GAUNTLET_URL = "https://www.dropbox.com/scl/fi/u3bjyjlb474tskbjyzmpg/gauntlet_w_ref_summaries.parquet?rlkey=qjsz6htflg77monh2y5jb3kya&dl=1"
        predict_dataset = Dataset.from_pandas(
            pd.read_parquet(GAUNTLET_URL)
        ).rename_columns(
            {
                "document_text": data_args.text_column,
            }
        )
        if data_args.summary_column != "summary":
            predict_dataset = predict_dataset.rename_columns(
                {
                    "summary": data_args.summary_column,
                }
            )
        if data_args.max_predict_samples is not None:
            max_predict_samples = min(
                len(predict_dataset), data_args.max_predict_samples
            )
            predict_dataset = predict_dataset.select(range(max_predict_samples))
        with training_args.main_process_first(
            desc="prediction dataset map pre-processing"
        ):
            predict_dataset = predict_dataset.map(
                preprocess_function,
                batched=True,
                batch_size=10,
                num_proc=data_args.preprocessing_num_workers,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on prediction dataset",
            )

    if data_args.do_predict_testset:
        max_target_length = data_args.val_max_target_length
        predict_dataset = raw_datasets["test"]
        if data_args.max_predict_samples is not None:
            max_predict_samples = min(
                len(predict_dataset), data_args.max_predict_samples
            )
            predict_dataset = predict_dataset.select(range(max_predict_samples))
        with training_args.main_process_first(
            desc="prediction dataset map pre-processing"
        ):
            predict_dataset = predict_dataset.map(
                preprocess_function,
                batched=True,
                batch_size=128 if data_args.preprocessing_num_workers > 30 else 512,
                num_proc=data_args.preprocessing_num_workers,
                remove_columns=column_names,
                load_from_cache_file=not data_args.overwrite_cache,
                desc="Running tokenizer on prediction dataset",
            )

    # Data collator
    label_pad_token_id = (
        -100 if data_args.ignore_pad_token_for_loss else tokenizer.pad_token_id
    )
    data_collator = DataCollatorForSeq2Seq(
        tokenizer,
        model=model,
        label_pad_token_id=label_pad_token_id,
        pad_to_multiple_of=8 if training_args.fp16 else None,
    )

    metric = evaluate.load("rouge")

    def postprocess_text(preds, labels):
        preds = [pred.strip() for pred in preds]
        labels = [label.strip() for label in labels]

        # rougeLSum expects newline after each sentence
        preds = ["\n".join(nltk.sent_tokenize(pred)) for pred in preds]
        labels = ["\n".join(nltk.sent_tokenize(label)) for label in labels]

        return preds, labels

    def compute_metrics(eval_preds):
        preds, labels = eval_preds
        if isinstance(preds, tuple):
            preds = preds[0]
        # Replace -100s used for padding as we can't decode them
        preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
        decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
        labels = np.where(labels != -100, labels, tokenizer.pad_token_id)
        decoded_labels = tokenizer.batch_decode(labels, skip_special_tokens=True)

        # Some simple post-processing
        decoded_preds, decoded_labels = postprocess_text(decoded_preds, decoded_labels)

        result = metric.compute(
            predictions=decoded_preds, references=decoded_labels, use_stemmer=True
        )
        result = {k: round(v * 100, 4) for k, v in result.items()}
        prediction_lens = [
            np.count_nonzero(pred != tokenizer.pad_token_id) for pred in preds
        ]
        result["gen_len"] = np.mean(prediction_lens)
        return result

    # Override the decoding parameters of Seq2SeqTrainer
    training_args.generation_max_length = (
        training_args.generation_max_length
        if training_args.generation_max_length is not None
        else data_args.val_max_target_length
    )
    training_args.generation_num_beams = (
        data_args.num_beams
        if data_args.num_beams is not None
        else training_args.generation_num_beams
    )

    if data_args.shuffle:
        logger.info(f"Shuffling train dataset with seed {training_args.seed}")
        train_dataset = train_dataset.shuffle(seed=training_args.seed)

    # Initialize our Trainer. He is very nice.
    trainer = Seq2SeqTrainer(
        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=data_collator,
        compute_metrics=(
            compute_metrics if training_args.predict_with_generate 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

        if model_args.disable_sdp_attention:
            train_result = trainer.train(resume_from_checkpoint=checkpoint)
        else:
            logger.info("Using sdp kernel for training")
            with sdpa_kernel(SDPBackend.FLASH_ATTENTION):
                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
    results = {}
    if training_args.do_eval:
        logger.info("*** Evaluate ***")
        metrics = trainer.evaluate(metric_key_prefix="eval")
        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))

        trainer.log_metrics("eval", metrics)
        trainer.save_metrics("eval", metrics)

    if training_args.do_predict:
        logger.info("*** Predict ***")

        predict_results = trainer.predict(predict_dataset, metric_key_prefix="predict")
        metrics = predict_results.metrics
        max_predict_samples = (
            data_args.max_predict_samples
            if data_args.max_predict_samples is not None
            else len(predict_dataset)
        )
        metrics["predict_samples"] = min(max_predict_samples, len(predict_dataset))

        trainer.log_metrics("predict", metrics)
        trainer.save_metrics("predict", metrics)

        if trainer.is_world_process_zero():
            if training_args.predict_with_generate:
                predictions = predict_results.predictions
                predictions = np.where(
                    predictions != -100, predictions, tokenizer.pad_token_id
                )
                predictions = tokenizer.batch_decode(
                    predictions,
                    skip_special_tokens=True,
                    clean_up_tokenization_spaces=True,
                )
                predictions = [pred.strip() for pred in predictions]
                output_prediction_file = os.path.join(
                    training_args.output_dir, "generated_gauntlet_predictions.txt"
                )
                with open(output_prediction_file, "w") as writer:
                    writer.write("\n".join(predictions))

    kwargs = {
        "finetuned_from": model_args.model_name_or_path,
        "tasks": "summarization",
        "dataset": data_args.dataset_name,
    }

    kwargs["language"] = data_args.lang if data_args.lang is not None else "en"

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

    logger.info("We are done here.")


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


if __name__ == "__main__":
    main()

t5_v1_1_samsum.sh

# Set environment variables
export WANDB_PROJECT="pileT5-summ"
export WANDB_WATCH="gradients"

NUM_WORKERS=$(lscpu -p | egrep -v '^#' | sort -u -t, -k 2,4 | wc -l)
echo "Number of CPU cores: $NUM_WORKERS"

# Set model ID
MODEL_ID="google/t5-v1_1-large"
MODEL_NAME=$(basename "$MODEL_ID")

# Run the summarization script
python ./run_summarization.py \
  --model_name_or_path $MODEL_ID \
  --do_train \
  --do_eval \
  --do_predict \
  --evaluation_strategy epoch \
  --dataset_name samsum \
  --text_column dialogue \
  --summary_column summary \
  --bf16 \
  --bf16_full_eval False \
  --data_seed 16919 \
  --dataloader_num_workers $NUM_WORKERS \
  --preprocessing_num_workers $NUM_WORKERS \
  --generation_max_length 512 \
  --gradient_accumulation_steps 16 \
  --gradient_checkpointing True \
  --hub_model_id BEE-spoke-data/${MODEL_NAME}-samsum \
  --hub_private_repo True \
  --hub_strategy "every_save" \
  --learning_rate 1e-4 \
  --load_best_model_at_end True \
  --logging_dir ./runtime/$MODEL_NAME-samsum-r1/logs \
  --logging_steps 3 \
  --lr_scheduler_type cosine \
  --max_eval_samples 300 \
  --max_grad_norm 1.0 \
  --max_source_length 1024 \
  --max_target_length 512 \
  --metric_for_best_model rouge2 \
  --num_beams 1 \
  --num_train_epochs 5 \
  --optim adamw_torch \
  --output_dir ./runtime/$MODEL_NAME-samsum-r1 \
  --overwrite_output_dir True \
  --pad_to_max_length False \
  --per_device_eval_batch_size 16 \
  --per_device_train_batch_size 8 \
  --predict_with_generate True \
  --push_to_hub \
  --report_to wandb \
  --run_name $MODEL_NAME-samsum-r1 \
  --save_strategy epoch \
  --seed 17868 \
  --sortish_sampler True \
  --tf32 True \
  --torch_compile_backend "inductor" \
  --val_max_target_length 512 \
  --warmup_ratio 0.05 \
  --weight_decay 0.01 \
  --greater_is_better True

text2text_instruct.sh

#!/bin/bash

# Set environment variables
export WANDB_PROJECT="text2text-instruct"
export WANDB_WATCH="gradients"
export WANDB_ENTITY="pszemraj"
export TOKENIZERS_PARALLELISM=true

NUM_WORKERS=$(lscpu -p | egrep -v '^#' | sort -u -t, -k 2,4 | wc -l)
echo "Number of CPU cores: $NUM_WORKERS"

# Set variables
MODEL_NAME_OR_PATH="pszemraj/tFINE-base-300m-SNI"
DATASET_NAME="pszemraj/infinity-instruct-7m-T2T_en"
DATASET_CONFIG="deduped-L1"
TEXT_COLUMN="instruction"
SUMMARY_COLUMN="response"
NUM_TRAIN_EPOCHS=1

LEARNING_RATE=1e-4
WARMUP_RATIO=0.05
OPTIM="adamw_torch_fused" # paged_adamw_32bit
LR_SCHEDULER_TYPE="cosine"
WEIGHT_DECAY=0.04
MAX_GRAD_NORM=1.0

PER_DEVICE_TRAIN_BATCH_SIZE=4
PER_DEVICE_EVAL_BATCH_SIZE=8
GRADIENT_ACCUMULATION_STEPS=16

MAX_SOURCE_LENGTH=1024
MAX_TARGET_LENGTH=1024
VAL_MAX_TARGET_LENGTH=1024
GENERATION_MAX_LENGTH=1024
NUM_BEAMS=1

SEED=17868
DATA_SEED=16919
MODEL_NAME=$(basename "$MODEL_NAME_OR_PATH")
DS_BASENAME=$(basename "$DATASET_NAME")
RUN_NAME="$MODEL_NAME-$DS_BASENAME-$MAX_SOURCE_LENGTH"
HUB_MODEL_ID="pszemraj/$RUN_NAME"
OUTPUT_DIR="./outputs/$RUN_NAME"

EVAL_STRATEGY="steps"
EVAL_STEPS=2000
MAX_EVAL_SAMPLES=150
LOGGING_DIR="./$OUTPUT_DIR/logs"
LOGGING_STEPS=10
SAVE_STRATEGY="steps"

python ./run_summarization.py \
  --model_name_or_path "$MODEL_NAME_OR_PATH" \
  --do_train \
  --do_eval \
  --evaluation_strategy "$EVAL_STRATEGY" --eval_steps "$EVAL_STEPS" \
  --dataset_name "$DATASET_NAME" --dataset_config_name "$DATASET_CONFIG" \
  --text_column "$TEXT_COLUMN" \
  --summary_column "$SUMMARY_COLUMN" \
  --bf16 \
  --bf16_full_eval False \
  --dataloader_num_workers "$NUM_WORKERS" \
  --filter True \
  --generation_max_length "$GENERATION_MAX_LENGTH" \
  --gradient_accumulation_steps "$GRADIENT_ACCUMULATION_STEPS" \
  --gradient_checkpointing True \
  --hub_model_id "$HUB_MODEL_ID" \
  --hub_private_repo True \
  --hub_strategy every_save \
  --include_num_input_tokens_seen True \
  --learning_rate "$LEARNING_RATE" \
  --logging_dir "$LOGGING_DIR" \
  --logging_steps "$LOGGING_STEPS" \
  --lr_scheduler_type "$LR_SCHEDULER_TYPE" \
  --max_eval_samples "$MAX_EVAL_SAMPLES" \
  --max_grad_norm "$MAX_GRAD_NORM" \
  --max_source_length "$MAX_SOURCE_LENGTH" \
  --max_target_length "$MAX_TARGET_LENGTH" \
  --num_beams "$NUM_BEAMS" \
  --num_train_epochs "$NUM_TRAIN_EPOCHS" \
  --optim "$OPTIM" \
  --output_dir "$OUTPUT_DIR" \
  --overwrite_output_dir True \
  --pad_to_max_length False \
  --per_device_eval_batch_size "$PER_DEVICE_EVAL_BATCH_SIZE" \
  --per_device_train_batch_size "$PER_DEVICE_TRAIN_BATCH_SIZE" \
  --predict_with_generate True \
  --preprocessing_num_workers "$NUM_WORKERS" \
  --push_to_hub \
  --report_to wandb \
  --run_name "$RUN_NAME" \
  --save_total_limit 1 \
  --seed "$SEED" \
  --sortish_sampler True \
  --tf32 True \
  --val_max_target_length "$VAL_MAX_TARGET_LENGTH" \
  --warmup_ratio "$WARMUP_RATIO" \
  --weight_decay "$WEIGHT_DECAY" \
  --torch_compile_backend "inductor"


# --use_fast_tokenizer False