快速开始

备注

在运行下述示例之前，需要您已经安装了PyTorch-NPU环境，有关环境安装，请参考 安装指南

一般来说，要在代码中使用NPU进行训练推理，需要做以下更改：

1. 导入torch_npu扩展包 import torch_npu

2. 将模型，以及模型输入上传到NPU上

device= torch.device("npu")
model = model.to(device)
input = input.to(device)

下面的实例演示了如何使用NPU进行训练和推理任务：

1. 单卡训练

以下代码使用了cifar10数据集在NPU上训练模型（截取自 PyTorch tutorials），请关注高亮的内容。

"""
Training an image classifier
----------------------------

We will do the following steps in order:

1. Load and normalize the CIFAR10 training and test datasets using
``torchvision``
1. Define a Convolutional Neural Network
2. Define a loss function
3. Train the network on the training data
4. Test the network on the test data

5. Load and normalize CIFAR10
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

Using ``torchvision``, it’s extremely easy to load CIFAR10.
"""
import torch
# 引入torch-npu包
import torch_npu

# 定义device
device = torch.device('npu:0' if torch.npu.is_available() else 'cpu')
print(device)

import torchvision
import torchvision.transforms as transforms

########################################################################
# The output of torchvision datasets are PILImage images of range [0, 1].
# We transform them to Tensors of normalized range [-1, 1].
transform = transforms.Compose(
    [transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])

batch_size = 4

trainset = torchvision.datasets.CIFAR10(root='./data', train=True,
                                        download=True, transform=transform)
trainloader = torch.utils.data.DataLoader(trainset, batch_size=batch_size,
                                        shuffle=True, num_workers=2)

testset = torchvision.datasets.CIFAR10(root='./data', train=False,
                                    download=True, transform=transform)
testloader = torch.utils.data.DataLoader(testset, batch_size=batch_size,
                                        shuffle=False, num_workers=2)

classes = ('plane', 'car', 'bird', 'cat',
        'deer', 'dog', 'frog', 'horse', 'ship', 'truck')

########################################################################
# 2. Define a Convolutional Neural Network
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
# Copy the neural network from the Neural Networks section before and modify it to
# take 3-channel images (instead of 1-channel images as it was defined).
import torch.nn as nn
import torch.nn.functional as F


class Net(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = torch.flatten(x, 1) # flatten all dimensions except batch
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

net = Net()

# 将模型加载到NPU上
net.to(device)

########################################################################
# 3. Define a Loss function and optimizer
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
# Let's use a Classification Cross-Entropy loss and SGD with momentum.
import torch.optim as optim

criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

########################################################################
# 4. Train the network
# ^^^^^^^^^^^^^^^^^^^^
#
# This is when things start to get interesting.
# We simply have to loop over our data iterator, and feed the inputs to the
# network and optimize.

for epoch in range(2):  # loop over the dataset multiple times

    running_loss = 0.0
    for i, data in enumerate(trainloader, 0):
        # get the inputs; data is a list of [inputs, labels]
        # 将input数据发送到NPU上
        inputs, labels = data[0].to(device), data[1].to(device)

        # zero the parameter gradients
        optimizer.zero_grad()

        # forward + backward + optimize
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        # print statistics
        running_loss += loss.item()
        if i % 2000 == 1999:    # print every 2000 mini-batches
            print(f'[{epoch + 1}, {i + 1:5d}] loss: {running_loss / 2000:.3f}')
            running_loss = 0.0

print('Finished Training')

########################################################################
# 5. Test the network on the test data
# ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
#
# We have trained the network for 2 passes over the training dataset.
# But we need to check if the network has learnt anything at all.
#
# We will check this by predicting the class label that the neural network
# outputs, and checking it against the ground-truth. If the prediction is
# correct, we add the sample to the list of correct predictions.
#
# Let us look at how the network performs on the whole dataset.
correct = 0
total = 0
# since we're not training, we don't need to calculate the gradients for our outputs
with torch.no_grad():
    for data in testloader:
        # 将input数据发送到NPU上
        images, labels = data[0].to(device), data[1].to(device)
        # calculate outputs by running images through the network
        outputs = net(images)
        # the class with the highest energy is what we choose as prediction
        _, predicted = torch.max(outputs.data, 1)
        total += labels.size(0)
        correct += (predicted == labels).sum().item()

print(f'Accuracy of the network on the 10000 test images: {100 * correct // total} %')
########################################################################
# That looks way better than chance, which is 10% accuracy (randomly picking
# a class out of 10 classes).
# Seems like the network learnt something.
#
# Hmmm, what are the classes that performed well, and the classes that did
# not perform well:

# prepare to count predictions for each class
correct_pred = {classname: 0 for classname in classes}
total_pred = {classname: 0 for classname in classes}

# again no gradients needed
with torch.no_grad():
    for data in testloader:
        # 将input数据发送到NPU上
        images, labels = data[0].to(device), data[1].to(device)
        outputs = net(images)
        _, predictions = torch.max(outputs, 1)
        # collect the correct predictions for each class
        for label, prediction in zip(labels, predictions):
            if label == prediction:
                correct_pred[classes[label]] += 1
            total_pred[classes[label]] += 1


# print accuracy for each class
for classname, correct_count in correct_pred.items():
    accuracy = 100 * float(correct_count) / total_pred[classname]
    print(f'Accuracy for class: {classname:5s} is {accuracy:.1f} %')

2. 使用DeepSpeed多卡并行训练

以下代码使用了cifar10数据集，使用DeepSpeed训练模型在多张NPU卡上进行模型训练（来自 DeepSpeed Examples），自DeepSpeed v0.12.6之后，代码无需任何修改，即可自动检测NPU并进行训练。

import argparse
import os

import deepspeed
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
from torchvision import transforms
from deepspeed.accelerator import get_accelerator
from deepspeed.moe.utils import split_params_into_different_moe_groups_for_optimizer


def add_argument():
    parser = argparse.ArgumentParser(description="CIFAR")

    # For train.
    parser.add_argument(
        "-e",
        "--epochs",
        default=30,
        type=int,
        help="number of total epochs (default: 30)",
    )
    parser.add_argument(
        "--local_rank",
        type=int,
        default=-1,
        help="local rank passed from distributed launcher",
    )
    parser.add_argument(
        "--log-interval",
        type=int,
        default=2000,
        help="output logging information at a given interval",
    )

    # For mixed precision training.
    parser.add_argument(
        "--dtype",
        default="fp16",
        type=str,
        choices=["bf16", "fp16", "fp32"],
        help="Datatype used for training",
    )

    # For ZeRO Optimization.
    parser.add_argument(
        "--stage",
        default=0,
        type=int,
        choices=[0, 1, 2, 3],
        help="Datatype used for training",
    )

    # For MoE (Mixture of Experts).
    parser.add_argument(
        "--moe",
        default=False,
        action="store_true",
        help="use deepspeed mixture of experts (moe)",
    )
    parser.add_argument(
        "--ep-world-size", default=1, type=int, help="(moe) expert parallel world size"
    )
    parser.add_argument(
        "--num-experts",
        type=int,
        nargs="+",
        default=[
            1,
        ],
        help="number of experts list, MoE related.",
    )
    parser.add_argument(
        "--mlp-type",
        type=str,
        default="standard",
        help="Only applicable when num-experts > 1, accepts [standard, residual]",
    )
    parser.add_argument(
        "--top-k", default=1, type=int, help="(moe) gating top 1 and 2 supported"
    )
    parser.add_argument(
        "--min-capacity",
        default=0,
        type=int,
        help="(moe) minimum capacity of an expert regardless of the capacity_factor",
    )
    parser.add_argument(
        "--noisy-gate-policy",
        default=None,
        type=str,
        help="(moe) noisy gating (only supported with top-1). Valid values are None, RSample, and Jitter",
    )
    parser.add_argument(
        "--moe-param-group",
        default=False,
        action="store_true",
        help="(moe) create separate moe param groups, required when using ZeRO w. MoE",
    )

    # Include DeepSpeed configuration arguments.
    parser = deepspeed.add_config_arguments(parser)

    args = parser.parse_args()

    return args


def create_moe_param_groups(model):
    """Create separate parameter groups for each expert."""
    parameters = {"params": [p for p in model.parameters()], "name": "parameters"}
    return split_params_into_different_moe_groups_for_optimizer(parameters)


def get_ds_config(args):
    """Get the DeepSpeed configuration dictionary."""
    ds_config = {
        "train_batch_size": 16,
        "steps_per_print": 2000,
        "optimizer": {
            "type": "Adam",
            "params": {
                "lr": 0.001,
                "betas": [0.8, 0.999],
                "eps": 1e-8,
                "weight_decay": 3e-7,
            },
        },
        "scheduler": {
            "type": "WarmupLR",
            "params": {
                "warmup_min_lr": 0,
                "warmup_max_lr": 0.001,
                "warmup_num_steps": 1000,
            },
        },
        "gradient_clipping": 1.0,
        "prescale_gradients": False,
        "bf16": {"enabled": args.dtype == "bf16"},
        "fp16": {
            "enabled": args.dtype == "fp16",
            "fp16_master_weights_and_grads": False,
            "loss_scale": 0,
            "loss_scale_window": 500,
            "hysteresis": 2,
            "min_loss_scale": 1,
            "initial_scale_power": 15,
        },
        "wall_clock_breakdown": False,
        "zero_optimization": {
            "stage": args.stage,
            "allgather_partitions": True,
            "reduce_scatter": True,
            "allgather_bucket_size": 50000000,
            "reduce_bucket_size": 50000000,
            "overlap_comm": True,
            "contiguous_gradients": True,
            "cpu_offload": False,
        },
    }
    return ds_config


class Net(nn.Module):
    def __init__(self, args):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.moe = args.moe
        if self.moe:
            fc3 = nn.Linear(84, 84)
            self.moe_layer_list = []
            for n_e in args.num_experts:
                # Create moe layers based on the number of experts.
                self.moe_layer_list.append(
                    deepspeed.moe.layer.MoE(
                        hidden_size=84,
                        expert=fc3,
                        num_experts=n_e,
                        ep_size=args.ep_world_size,
                        use_residual=args.mlp_type == "residual",
                        k=args.top_k,
                        min_capacity=args.min_capacity,
                        noisy_gate_policy=args.noisy_gate_policy,
                    )
                )
            self.moe_layer_list = nn.ModuleList(self.moe_layer_list)
            self.fc4 = nn.Linear(84, 10)
        else:
            self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        if self.moe:
            for layer in self.moe_layer_list:
                x, _, _ = layer(x)
            x = self.fc4(x)
        else:
            x = self.fc3(x)
        return x


def test(model_engine, testset, local_device, target_dtype, test_batch_size=4):
    """Test the network on the test data.

    Args:
        model_engine (deepspeed.runtime.engine.DeepSpeedEngine): the DeepSpeed engine.
        testset (torch.utils.data.Dataset): the test dataset.
        local_device (str): the local device name.
        target_dtype (torch.dtype): the target datatype for the test data.
        test_batch_size (int): the test batch size.

    """
    # The 10 classes for CIFAR10.
    classes = (
        "plane",
        "car",
        "bird",
        "cat",
        "deer",
        "dog",
        "frog",
        "horse",
        "ship",
        "truck",
    )

    # Define the test dataloader.
    testloader = torch.utils.data.DataLoader(
        testset, batch_size=test_batch_size, shuffle=False, num_workers=0
    )

    # For total accuracy.
    correct, total = 0, 0
    # For accuracy per class.
    class_correct = list(0.0 for i in range(10))
    class_total = list(0.0 for i in range(10))

    # Start testing.
    model_engine.eval()
    with torch.no_grad():
        for data in testloader:
            images, labels = data
            if target_dtype != None:
                images = images.to(target_dtype)
            outputs = model_engine(images.to(local_device))
            _, predicted = torch.max(outputs.data, 1)
            # Count the total accuracy.
            total += labels.size(0)
            correct += (predicted == labels.to(local_device)).sum().item()

            # Count the accuracy per class.
            batch_correct = (predicted == labels.to(local_device)).squeeze()
            for i in range(test_batch_size):
                label = labels[i]
                class_correct[label] += batch_correct[i].item()
                class_total[label] += 1

    if model_engine.local_rank == 0:
        print(
            f"Accuracy of the network on the {total} test images: {100 * correct / total : .0f} %"
        )

        # For all classes, print the accuracy.
        for i in range(10):
            print(
                f"Accuracy of {classes[i] : >5s} : {100 * class_correct[i] / class_total[i] : 2.0f} %"
            )


def main(args):
    # Initialize DeepSpeed distributed backend.
    deepspeed.init_distributed()
    _local_rank = int(os.environ.get("LOCAL_RANK"))
    get_accelerator().set_device(_local_rank)

    ########################################################################
    # Step1. Data Preparation.
    #
    # The output of torchvision datasets are PILImage images of range [0, 1].
    # We transform them to Tensors of normalized range [-1, 1].
    #
    # Note:
    #     If running on Windows and you get a BrokenPipeError, try setting
    #     the num_worker of torch.utils.data.DataLoader() to 0.
    ########################################################################
    transform = transforms.Compose(
        [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]
    )

    if torch.distributed.get_rank() != 0:
        # Might be downloading cifar data, let rank 0 download first.
        torch.distributed.barrier()

    # Load or download cifar data.
    trainset = torchvision.datasets.CIFAR10(
        root="./data", train=True, download=True, transform=transform
    )
    testset = torchvision.datasets.CIFAR10(
        root="./data", train=False, download=True, transform=transform
    )

    if torch.distributed.get_rank() == 0:
        # Cifar data is downloaded, indicate other ranks can proceed.
        torch.distributed.barrier()

    ########################################################################
    # Step 2. Define the network with DeepSpeed.
    #
    # First, we define a Convolution Neural Network.
    # Then, we define the DeepSpeed configuration dictionary and use it to
    # initialize the DeepSpeed engine.
    ########################################################################
    net = Net(args)

    # Get list of parameters that require gradients.
    parameters = filter(lambda p: p.requires_grad, net.parameters())

    # If using MoE, create separate param groups for each expert.
    if args.moe_param_group:
        parameters = create_moe_param_groups(net)

    # Initialize DeepSpeed to use the following features.
    #   1) Distributed model.
    #   2) Distributed data loader.
    #   3) DeepSpeed optimizer.
    ds_config = get_ds_config(args)
    model_engine, optimizer, trainloader, __ = deepspeed.initialize(
        args=args,
        model=net,
        model_parameters=parameters,
        training_data=trainset,
        config=ds_config,
    )

    # Get the local device name (str) and local rank (int).
    local_device = get_accelerator().device_name(model_engine.local_rank)
    local_rank = model_engine.local_rank

    # For float32, target_dtype will be None so no datatype conversion needed.
    target_dtype = None
    if model_engine.bfloat16_enabled():
        target_dtype = torch.bfloat16
    elif model_engine.fp16_enabled():
        target_dtype = torch.half

    # Define the Classification Cross-Entropy loss function.
    criterion = nn.CrossEntropyLoss()

    ########################################################################
    # Step 3. Train the network.
    #
    # This is when things start to get interesting.
    # We simply have to loop over our data iterator, and feed the inputs to the
    # network and optimize. (DeepSpeed handles the distributed details for us!)
    ########################################################################

    for epoch in range(args.epochs):  # loop over the dataset multiple times
        running_loss = 0.0
        for i, data in enumerate(trainloader):
            # Get the inputs. ``data`` is a list of [inputs, labels].
            inputs, labels = data[0].to(local_device), data[1].to(local_device)

            # Try to convert to target_dtype if needed.
            if target_dtype != None:
                inputs = inputs.to(target_dtype)

            outputs = model_engine(inputs)
            loss = criterion(outputs, labels)

            model_engine.backward(loss)
            model_engine.step()

            # Print statistics
            running_loss += loss.item()
            if local_rank == 0 and i % args.log_interval == (
                args.log_interval - 1
            ):  # Print every log_interval mini-batches.
                print(
                    f"[{epoch + 1 : d}, {i + 1 : 5d}] loss: {running_loss / args.log_interval : .3f}"
                )
                running_loss = 0.0
    print("Finished Training")

    ########################################################################
    # Step 4. Test the network on the test data.
    ########################################################################
    test(model_engine, testset, local_device, target_dtype)


if __name__ == "__main__":
    args = add_argument()
    main(args)

3. 使用Transforms进行模型微调

以下代码使用了Transforms对LLM进行微调（来自 transforms examples），自transforms xxx版本以及accelerator 0.21.0版本以后，代码无需任何修改，即可自动检测NPU并进行。

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

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

import logging
import math
import os
import sys
from dataclasses import dataclass, field
from itertools import chain
from typing import Optional

import datasets
import evaluate
import torch
from datasets import load_dataset

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("4.52.0.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.
    """

    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."
            )
        },
    )

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


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

    dataset_name: Optional[str] = field(
        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
    )
    dataset_config_name: Optional[str] = field(
        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
    )
    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."


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

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

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

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

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

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

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

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

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

    # Get the datasets: you can either provide your own CSV/JSON/TXT training and evaluation files (see below)
    # or just provide the name of one of the public datasets available on the hub at https://huggingface.co/datasets/
    # (the dataset will be downloaded automatically from the datasets Hub).
    #
    # For CSV/JSON files, this script will use the column called 'text' or the first column if no column called
    # 'text' is found. You can easily tweak this behavior (see below).
    #
    # In distributed training, the load_dataset function guarantee that only one local process can concurrently
    # download the dataset.
    if data_args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        raw_datasets = load_dataset(
            data_args.dataset_name,
            data_args.dataset_config_name,
            cache_dir=model_args.cache_dir,
            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

    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,
            )

    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))

    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))

        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 __name__ == "__main__":
    main()

python run_clm.py \
    --model_name_or_path openai-community/gpt2 \
    --train_file path_to_train_file \
    --validation_file path_to_validation_file \
    --per_device_train_batch_size 8 \
    --per_device_eval_batch_size 8 \
    --do_train \
    --do_eval \
    --output_dir /tmp/test-clm

4. 使用Diffusers进行模型微调

以下代码使用了Diffusers对文生图模型进行微调（来自 diffusers examples），自diffusers v0.27.0版本以后，代码无需任何修改，即可自动检测NPU并进行。

#!/usr/bin/env python
# coding=utf-8
# Copyright 2025 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.

import argparse
import logging
import math
import os
import random
import shutil
from contextlib import nullcontext
from pathlib import Path

import accelerate
import datasets
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.state import AcceleratorState
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer
from transformers.utils import ContextManagers

import diffusers
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel
from diffusers.optimization import get_scheduler
from diffusers.training_utils import EMAModel, compute_dream_and_update_latents, compute_snr
from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_xformers_available
from diffusers.utils.torch_utils import is_compiled_module


if is_wandb_available():
    import wandb


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.34.0.dev0")

logger = get_logger(__name__, log_level="INFO")

DATASET_NAME_MAPPING = {
    "lambdalabs/naruto-blip-captions": ("image", "text"),
}


def save_model_card(
    args,
    repo_id: str,
    images: list = None,
    repo_folder: str = None,
):
    img_str = ""
    if len(images) > 0:
        image_grid = make_image_grid(images, 1, len(args.validation_prompts))
        image_grid.save(os.path.join(repo_folder, "val_imgs_grid.png"))
        img_str += "![val_imgs_grid](./val_imgs_grid.png)\n"

    model_description = f"""
# Text-to-image finetuning - {repo_id}

This pipeline was finetuned from **{args.pretrained_model_name_or_path}** on the **{args.dataset_name}** dataset. Below are some example images generated with the finetuned pipeline using the following prompts: {args.validation_prompts}: \n
{img_str}

## Pipeline usage

You can use the pipeline like so:

```python
from diffusers import DiffusionPipeline
import torch

pipeline = DiffusionPipeline.from_pretrained("{repo_id}", torch_dtype=torch.float16)
prompt = "{args.validation_prompts[0]}"
image = pipeline(prompt).images[0]
image.save("my_image.png")
```

## Training info

These are the key hyperparameters used during training:

* Epochs: {args.num_train_epochs}
* Learning rate: {args.learning_rate}
* Batch size: {args.train_batch_size}
* Gradient accumulation steps: {args.gradient_accumulation_steps}
* Image resolution: {args.resolution}
* Mixed-precision: {args.mixed_precision}

"""
    wandb_info = ""
    if is_wandb_available():
        wandb_run_url = None
        if wandb.run is not None:
            wandb_run_url = wandb.run.url

    if wandb_run_url is not None:
        wandb_info = f"""
More information on all the CLI arguments and the environment are available on your [`wandb` run page]({wandb_run_url}).
"""

    model_description += wandb_info

    model_card = load_or_create_model_card(
        repo_id_or_path=repo_id,
        from_training=True,
        license="creativeml-openrail-m",
        base_model=args.pretrained_model_name_or_path,
        model_description=model_description,
        inference=True,
    )

    tags = ["stable-diffusion", "stable-diffusion-diffusers", "text-to-image", "diffusers", "diffusers-training"]
    model_card = populate_model_card(model_card, tags=tags)

    model_card.save(os.path.join(repo_folder, "README.md"))


def log_validation(vae, text_encoder, tokenizer, unet, args, accelerator, weight_dtype, epoch):
    logger.info("Running validation... ")

    pipeline = StableDiffusionPipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        vae=accelerator.unwrap_model(vae),
        text_encoder=accelerator.unwrap_model(text_encoder),
        tokenizer=tokenizer,
        unet=accelerator.unwrap_model(unet),
        safety_checker=None,
        revision=args.revision,
        variant=args.variant,
        torch_dtype=weight_dtype,
    )
    pipeline = pipeline.to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)

    if args.enable_xformers_memory_efficient_attention:
        pipeline.enable_xformers_memory_efficient_attention()

    if args.seed is None:
        generator = None
    else:
        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)

    images = []
    for i in range(len(args.validation_prompts)):
        if torch.backends.mps.is_available():
            autocast_ctx = nullcontext()
        else:
            autocast_ctx = torch.autocast(accelerator.device.type)

        with autocast_ctx:
            image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]

        images.append(image)

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            np_images = np.stack([np.asarray(img) for img in images])
            tracker.writer.add_images("validation", np_images, epoch, dataformats="NHWC")
        elif tracker.name == "wandb":
            tracker.log(
                {
                    "validation": [
                        wandb.Image(image, caption=f"{i}: {args.validation_prompts[i]}")
                        for i, image in enumerate(images)
                    ]
                }
            )
        else:
            logger.warning(f"image logging not implemented for {tracker.name}")

    del pipeline
    torch.cuda.empty_cache()

    return images


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
    )
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--variant",
        type=str,
        default=None,
        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
    )
    parser.add_argument(
        "--dataset_name",
        type=str,
        default=None,
        help=(
            "The name of the Dataset (from the HuggingFace hub) to train on (could be your own, possibly private,"
            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
            " or to a folder containing files that 🤗 Datasets can understand."
        ),
    )
    parser.add_argument(
        "--dataset_config_name",
        type=str,
        default=None,
        help="The config of the Dataset, leave as None if there's only one config.",
    )
    parser.add_argument(
        "--train_data_dir",
        type=str,
        default=None,
        help=(
            "A folder containing the training data. Folder contents must follow the structure described in"
            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
    )
    parser.add_argument(
        "--caption_column",
        type=str,
        default="text",
        help="The column of the dataset containing a caption or a list of captions.",
    )
    parser.add_argument(
        "--max_train_samples",
        type=int,
        default=None,
        help=(
            "For debugging purposes or quicker training, truncate the number of training examples to this "
            "value if set."
        ),
    )
    parser.add_argument(
        "--validation_prompts",
        type=str,
        default=None,
        nargs="+",
        help=("A set of prompts evaluated every `--validation_epochs` and logged to `--report_to`."),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="sd-model-finetuned",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--resolution",
        type=int,
        default=512,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--center_crop",
        default=False,
        action="store_true",
        help=(
            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
            " cropped. The images will be resized to the resolution first before cropping."
        ),
    )
    parser.add_argument(
        "--random_flip",
        action="store_true",
        help="whether to randomly flip images horizontally",
    )
    parser.add_argument(
        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument("--num_train_epochs", type=int, default=100)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="constant",
        help=(
            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
            ' "constant", "constant_with_warmup"]'
        ),
    )
    parser.add_argument(
        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument(
        "--snr_gamma",
        type=float,
        default=None,
        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
        "More details here: https://arxiv.org/abs/2303.09556.",
    )
    parser.add_argument(
        "--dream_training",
        action="store_true",
        help=(
            "Use the DREAM training method, which makes training more efficient and accurate at the "
            "expense of doing an extra forward pass. See: https://arxiv.org/abs/2312.00210"
        ),
    )
    parser.add_argument(
        "--dream_detail_preservation",
        type=float,
        default=1.0,
        help="Dream detail preservation factor p (should be greater than 0; default=1.0, as suggested in the paper)",
    )
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    parser.add_argument(
        "--allow_tf32",
        action="store_true",
        help=(
            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
        ),
    )
    parser.add_argument("--use_ema", action="store_true", help="Whether to use EMA model.")
    parser.add_argument("--offload_ema", action="store_true", help="Offload EMA model to CPU during training step.")
    parser.add_argument("--foreach_ema", action="store_true", help="Use faster foreach implementation of EMAModel.")
    parser.add_argument(
        "--non_ema_revision",
        type=str,
        default=None,
        required=False,
        help=(
            "Revision of pretrained non-ema model identifier. Must be a branch, tag or git identifier of the local or"
            " remote repository specified with --pretrained_model_name_or_path."
        ),
    )
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help=(
            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
        ),
    )
    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--prediction_type",
        type=str,
        default=None,
        help="The prediction_type that shall be used for training. Choose between 'epsilon' or 'v_prediction' or leave `None`. If left to `None` the default prediction type of the scheduler: `noise_scheduler.config.prediction_type` is chosen.",
    )
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16"],
        help=(
            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="tensorboard",
        help=(
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
        ),
    )
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints are only suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    parser.add_argument(
        "--checkpoints_total_limit",
        type=int,
        default=None,
        help=("Max number of checkpoints to store."),
    )
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help=(
            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
        ),
    )
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
    parser.add_argument("--noise_offset", type=float, default=0, help="The scale of noise offset.")
    parser.add_argument(
        "--validation_epochs",
        type=int,
        default=5,
        help="Run validation every X epochs.",
    )
    parser.add_argument(
        "--tracker_project_name",
        type=str,
        default="text2image-fine-tune",
        help=(
            "The `project_name` argument passed to Accelerator.init_trackers for"
            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
        ),
    )

    args = parser.parse_args()
    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    # Sanity checks
    if args.dataset_name is None and args.train_data_dir is None:
        raise ValueError("Need either a dataset name or a training folder.")

    # default to using the same revision for the non-ema model if not specified
    if args.non_ema_revision is None:
        args.non_ema_revision = args.revision

    return args


def main():
    args = parse_args()

    if args.report_to == "wandb" and args.hub_token is not None:
        raise ValueError(
            "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
            " Please use `huggingface-cli login` to authenticate with the Hub."
        )

    if args.non_ema_revision is not None:
        deprecate(
            "non_ema_revision!=None",
            "0.15.0",
            message=(
                "Downloading 'non_ema' weights from revision branches of the Hub is deprecated. Please make sure to"
                " use `--variant=non_ema` instead."
            ),
        )
    logging_dir = os.path.join(args.output_dir, args.logging_dir)

    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
    )

    # Disable AMP for MPS.
    if torch.backends.mps.is_available():
        accelerator.native_amp = False

    # 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_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        datasets.utils.logging.set_verbosity_error()
        transformers.utils.logging.set_verbosity_error()
        diffusers.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:
            os.makedirs(args.output_dir, exist_ok=True)

        if args.push_to_hub:
            repo_id = create_repo(
                repo_id=args.hub_model_id or Path(args.output_dir).name, exist_ok=True, token=args.hub_token
            ).repo_id

    # Load scheduler, tokenizer and models.
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler")
    tokenizer = CLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision
    )

    def deepspeed_zero_init_disabled_context_manager():
        """
        returns either a context list that includes one that will disable zero.Init or an empty context list
        """
        deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
        if deepspeed_plugin is None:
            return []

        return [deepspeed_plugin.zero3_init_context_manager(enable=False)]

    # Currently Accelerate doesn't know how to handle multiple models under Deepspeed ZeRO stage 3.
    # For this to work properly all models must be run through `accelerate.prepare`. But accelerate
    # will try to assign the same optimizer with the same weights to all models during
    # `deepspeed.initialize`, which of course doesn't work.
    #
    # For now the following workaround will partially support Deepspeed ZeRO-3, by excluding the 2
    # frozen models from being partitioned during `zero.Init` which gets called during
    # `from_pretrained` So CLIPTextModel and AutoencoderKL will not enjoy the parameter sharding
    # across multiple gpus and only UNet2DConditionModel will get ZeRO sharded.
    with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
        text_encoder = CLIPTextModel.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
        )
        vae = AutoencoderKL.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant
        )

    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.non_ema_revision
    )

    # Freeze vae and text_encoder and set unet to trainable
    vae.requires_grad_(False)
    text_encoder.requires_grad_(False)
    unet.train()

    # Create EMA for the unet.
    if args.use_ema:
        ema_unet = UNet2DConditionModel.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant
        )
        ema_unet = EMAModel(
            ema_unet.parameters(),
            model_cls=UNet2DConditionModel,
            model_config=ema_unet.config,
            foreach=args.foreach_ema,
        )

    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
                logger.warning(
                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
                )
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError("xformers is not available. Make sure it is installed correctly")

    # `accelerate` 0.16.0 will have better support for customized saving
    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
        # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
        def save_model_hook(models, weights, output_dir):
            if accelerator.is_main_process:
                if args.use_ema:
                    ema_unet.save_pretrained(os.path.join(output_dir, "unet_ema"))

                for i, model in enumerate(models):
                    model.save_pretrained(os.path.join(output_dir, "unet"))

                    # make sure to pop weight so that corresponding model is not saved again
                    weights.pop()

        def load_model_hook(models, input_dir):
            if args.use_ema:
                load_model = EMAModel.from_pretrained(
                    os.path.join(input_dir, "unet_ema"), UNet2DConditionModel, foreach=args.foreach_ema
                )
                ema_unet.load_state_dict(load_model.state_dict())
                if args.offload_ema:
                    ema_unet.pin_memory()
                else:
                    ema_unet.to(accelerator.device)
                del load_model

            for _ in range(len(models)):
                # pop models so that they are not loaded again
                model = models.pop()

                # load diffusers style into model
                load_model = UNet2DConditionModel.from_pretrained(input_dir, subfolder="unet")
                model.register_to_config(**load_model.config)

                model.load_state_dict(load_model.state_dict())
                del load_model

        accelerator.register_save_state_pre_hook(save_model_hook)
        accelerator.register_load_state_pre_hook(load_model_hook)

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if args.allow_tf32:
        torch.backends.cuda.matmul.allow_tf32 = True

    if args.scale_lr:
        args.learning_rate = (
            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
        )

    # Initialize the optimizer
    if args.use_8bit_adam:
        try:
            import bitsandbytes as bnb
        except ImportError:
            raise ImportError(
                "Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
            )

        optimizer_cls = bnb.optim.AdamW8bit
    else:
        optimizer_cls = torch.optim.AdamW

    optimizer = optimizer_cls(
        unet.parameters(),
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # Get the datasets: you can either provide your own training and evaluation files (see below)
    # or specify a Dataset from the hub (the dataset will be downloaded automatically from the datasets Hub).

    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            args.dataset_name,
            args.dataset_config_name,
            cache_dir=args.cache_dir,
            data_dir=args.train_data_dir,
        )
    else:
        data_files = {}
        if args.train_data_dir is not None:
            data_files["train"] = os.path.join(args.train_data_dir, "**")
        dataset = load_dataset(
            "imagefolder",
            data_files=data_files,
            cache_dir=args.cache_dir,
        )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder

    # Preprocessing the datasets.
    # We need to tokenize inputs and targets.
    column_names = dataset["train"].column_names

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

    # Preprocessing the datasets.
    # We need to tokenize input captions and transform the images.
    def tokenize_captions(examples, is_train=True):
        captions = []
        for caption in examples[caption_column]:
            if isinstance(caption, str):
                captions.append(caption)
            elif isinstance(caption, (list, np.ndarray)):
                # take a random caption if there are multiple
                captions.append(random.choice(caption) if is_train else caption[0])
            else:
                raise ValueError(
                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
                )
        inputs = tokenizer(
            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
        )
        return inputs.input_ids

    # Preprocessing the datasets.
    train_transforms = transforms.Compose(
        [
            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
            transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
            transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
            transforms.ToTensor(),
            transforms.Normalize([0.5], [0.5]),
        ]
    )

    def preprocess_train(examples):
        images = [image.convert("RGB") for image in examples[image_column]]
        examples["pixel_values"] = [train_transforms(image) for image in images]
        examples["input_ids"] = tokenize_captions(examples)
        return examples

    with accelerator.main_process_first():
        if args.max_train_samples is not None:
            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
        # Set the training transforms
        train_dataset = dataset["train"].with_transform(preprocess_train)

    def collate_fn(examples):
        pixel_values = torch.stack([example["pixel_values"] for example in examples])
        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
        input_ids = torch.stack([example["input_ids"] for example in examples])
        return {"pixel_values": pixel_values, "input_ids": input_ids}

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        num_workers=args.dataloader_num_workers,
    )

    # Scheduler and math around the number of training steps.
    # Check the PR https://github.com/huggingface/diffusers/pull/8312 for detailed explanation.
    num_warmup_steps_for_scheduler = args.lr_warmup_steps * accelerator.num_processes
    if args.max_train_steps is None:
        len_train_dataloader_after_sharding = math.ceil(len(train_dataloader) / accelerator.num_processes)
        num_update_steps_per_epoch = math.ceil(len_train_dataloader_after_sharding / args.gradient_accumulation_steps)
        num_training_steps_for_scheduler = (
            args.num_train_epochs * num_update_steps_per_epoch * accelerator.num_processes
        )
    else:
        num_training_steps_for_scheduler = args.max_train_steps * accelerator.num_processes

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=num_warmup_steps_for_scheduler,
        num_training_steps=num_training_steps_for_scheduler,
    )

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

    if args.use_ema:
        if args.offload_ema:
            ema_unet.pin_memory()
        else:
            ema_unet.to(accelerator.device)

    # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
        args.mixed_precision = accelerator.mixed_precision
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16
        args.mixed_precision = accelerator.mixed_precision

    # Move text_encode and vae to gpu and cast to weight_dtype
    text_encoder.to(accelerator.device, dtype=weight_dtype)
    vae.to(accelerator.device, dtype=weight_dtype)

    # 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 args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        if num_training_steps_for_scheduler != args.max_train_steps * accelerator.num_processes:
            logger.warning(
                f"The length of the 'train_dataloader' after 'accelerator.prepare' ({len(train_dataloader)}) does not match "
                f"the expected length ({len_train_dataloader_after_sharding}) when the learning rate scheduler was created. "
                f"This inconsistency may result in the learning rate scheduler not functioning properly."
            )
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        tracker_config = dict(vars(args))
        tracker_config.pop("validation_prompts")
        accelerator.init_trackers(args.tracker_project_name, tracker_config)

    # Function for unwrapping if model was compiled with `torch.compile`.
    def unwrap_model(model):
        model = accelerator.unwrap_model(model)
        model = model._orig_mod if is_compiled_module(model) else model
        return model

    # Train!
    total_batch_size = args.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.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}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the most recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            path = dirs[-1] if len(dirs) > 0 else None

        if path is None:
            accelerator.print(
                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            args.resume_from_checkpoint = None
            initial_global_step = 0
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch

    else:
        initial_global_step = 0

    progress_bar = tqdm(
        range(0, args.max_train_steps),
        initial=initial_global_step,
        desc="Steps",
        # Only show the progress bar once on each machine.
        disable=not accelerator.is_local_main_process,
    )

    for epoch in range(first_epoch, args.num_train_epochs):
        train_loss = 0.0
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                # Convert images to latent space
                latents = vae.encode(batch["pixel_values"].to(weight_dtype)).latent_dist.sample()
                latents = latents * vae.config.scaling_factor

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(latents)
                if args.noise_offset:
                    # https://www.crosslabs.org//blog/diffusion-with-offset-noise
                    noise += args.noise_offset * torch.randn(
                        (latents.shape[0], latents.shape[1], 1, 1), device=latents.device
                    )
                if args.input_perturbation:
                    new_noise = noise + args.input_perturbation * torch.randn_like(noise)
                bsz = latents.shape[0]
                # Sample a random timestep for each image
                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
                timesteps = timesteps.long()

                # Add noise to the latents according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                if args.input_perturbation:
                    noisy_latents = noise_scheduler.add_noise(latents, new_noise, timesteps)
                else:
                    noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

                # Get the text embedding for conditioning
                encoder_hidden_states = text_encoder(batch["input_ids"], return_dict=False)[0]

                # Get the target for loss depending on the prediction type
                if args.prediction_type is not None:
                    # set prediction_type of scheduler if defined
                    noise_scheduler.register_to_config(prediction_type=args.prediction_type)

                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(latents, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

                if args.dream_training:
                    noisy_latents, target = compute_dream_and_update_latents(
                        unet,
                        noise_scheduler,
                        timesteps,
                        noise,
                        noisy_latents,
                        target,
                        encoder_hidden_states,
                        args.dream_detail_preservation,
                    )

                # Predict the noise residual and compute loss
                model_pred = unet(noisy_latents, timesteps, encoder_hidden_states, return_dict=False)[0]

                if args.snr_gamma is None:
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
                else:
                    # Compute loss-weights as per Section 3.4 of https://arxiv.org/abs/2303.09556.
                    # Since we predict the noise instead of x_0, the original formulation is slightly changed.
                    # This is discussed in Section 4.2 of the same paper.
                    snr = compute_snr(noise_scheduler, timesteps)
                    mse_loss_weights = torch.stack([snr, args.snr_gamma * torch.ones_like(timesteps)], dim=1).min(
                        dim=1
                    )[0]
                    if noise_scheduler.config.prediction_type == "epsilon":
                        mse_loss_weights = mse_loss_weights / snr
                    elif noise_scheduler.config.prediction_type == "v_prediction":
                        mse_loss_weights = mse_loss_weights / (snr + 1)

                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="none")
                    loss = loss.mean(dim=list(range(1, len(loss.shape)))) * mse_loss_weights
                    loss = loss.mean()

                # Gather the losses across all processes for logging (if we use distributed training).
                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
                train_loss += avg_loss.item() / args.gradient_accumulation_steps

                # Backpropagate
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                if args.use_ema:
                    if args.offload_ema:
                        ema_unet.to(device="cuda", non_blocking=True)
                    ema_unet.step(unet.parameters())
                    if args.offload_ema:
                        ema_unet.to(device="cpu", non_blocking=True)
                progress_bar.update(1)
                global_step += 1
                accelerator.log({"train_loss": train_loss}, step=global_step)
                train_loss = 0.0

                if global_step % args.checkpointing_steps == 0:
                    if accelerator.is_main_process:
                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                        if args.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(args.output_dir)
                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))

                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
                            if len(checkpoints) >= args.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
                                removing_checkpoints = checkpoints[0:num_to_remove]

                                logger.info(
                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
                                )
                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")

                                for removing_checkpoint in removing_checkpoints:
                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)

                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")

            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
            progress_bar.set_postfix(**logs)

            if global_step >= args.max_train_steps:
                break

        if accelerator.is_main_process:
            if args.validation_prompts is not None and epoch % args.validation_epochs == 0:
                if args.use_ema:
                    # Store the UNet parameters temporarily and load the EMA parameters to perform inference.
                    ema_unet.store(unet.parameters())
                    ema_unet.copy_to(unet.parameters())
                log_validation(
                    vae,
                    text_encoder,
                    tokenizer,
                    unet,
                    args,
                    accelerator,
                    weight_dtype,
                    global_step,
                )
                if args.use_ema:
                    # Switch back to the original UNet parameters.
                    ema_unet.restore(unet.parameters())

    # Create the pipeline using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = unwrap_model(unet)
        if args.use_ema:
            ema_unet.copy_to(unet.parameters())

        pipeline = StableDiffusionPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            text_encoder=text_encoder,
            vae=vae,
            unet=unet,
            revision=args.revision,
            variant=args.variant,
        )
        pipeline.save_pretrained(args.output_dir)

        # Run a final round of inference.
        images = []
        if args.validation_prompts is not None:
            logger.info("Running inference for collecting generated images...")
            pipeline = pipeline.to(accelerator.device)
            pipeline.torch_dtype = weight_dtype
            pipeline.set_progress_bar_config(disable=True)

            if args.enable_xformers_memory_efficient_attention:
                pipeline.enable_xformers_memory_efficient_attention()

            if args.seed is None:
                generator = None
            else:
                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)

            for i in range(len(args.validation_prompts)):
                with torch.autocast("cuda"):
                    image = pipeline(args.validation_prompts[i], num_inference_steps=20, generator=generator).images[0]
                images.append(image)

        if args.push_to_hub:
            save_model_card(args, repo_id, images, repo_folder=args.output_dir)
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )

    accelerator.end_training()


if __name__ == "__main__":
    main()

export MODEL_NAME="CompVis/stable-diffusion-v1-4"
export DATASET_NAME="lambdalabs/naruto-blip-captions"

accelerate launch --mixed_precision="fp16"  train_text_to_image.py \
--pretrained_model_name_or_path=$MODEL_NAME \
--dataset_name=$DATASET_NAME \
--use_ema \
--resolution=512 --center_crop --random_flip \
--train_batch_size=1 \
--gradient_accumulation_steps=4 \
--gradient_checkpointing \
--max_train_steps=15000 \
--learning_rate=1e-05 \
--max_grad_norm=1 \
--lr_scheduler="constant" --lr_warmup_steps=0 \
--output_dir="sd-pokemon-model"