Führen Sie PyTorch-Trainingsjobs mit dem SageMaker Training Compiler aus

Um ein PyTorch-Modell zu kompilieren und zu trainieren, konfigurieren Sie einen PyTorch-Schätzer von SageMaker AI mit dem SageMaker Training Compiler, wie im folgenden Codebeispiel gezeigt.

from sagemaker.pytorch import PyTorch, TrainingCompilerConfig

# the original max batch size that can fit into GPU memory without compiler
batch_size_native=12
learning_rate_native=float('5e-5')

# an updated max batch size that can fit into GPU memory with compiler
batch_size=64

# update learning rate
learning_rate=learning_rate_native/batch_size_native*batch_size

hyperparameters={
    "n_gpus": 1,
    "batch_size": batch_size,
    "learning_rate": learning_rate
}

pytorch_estimator=PyTorch(
    entry_point='train.py',
    source_dir='path-to-requirements-file', # Optional. Add this if need to install additional packages.
    instance_count=1,
    instance_type='ml.p3.2xlarge',
    framework_version='1.13.1',
    py_version='py3',
    hyperparameters=hyperparameters,
    compiler_config=TrainingCompilerConfig(),
    disable_profiler=True,
    debugger_hook_config=False
)

pytorch_estimator.fit()

Um ein Transformer-Modell mit PyTorch zu kompilieren und zu trainieren, konfigurieren Sie einen Hugging-Face-Schätzer von SageMaker AI mit dem SageMaker Training Compiler, wie im folgenden Codebeispiel gezeigt.

from sagemaker.huggingface import HuggingFace, TrainingCompilerConfig

# the original max batch size that can fit into GPU memory without compiler
batch_size_native=12
learning_rate_native=float('5e-5')

# an updated max batch size that can fit into GPU memory with compiler
batch_size=64

# update learning rate
learning_rate=learning_rate_native/batch_size_native*batch_size

hyperparameters={
    "n_gpus": 1,
    "batch_size": batch_size,
    "learning_rate": learning_rate
}

pytorch_huggingface_estimator=HuggingFace(
    entry_point='train.py',
    instance_count=1,
    instance_type='ml.p3.2xlarge',
    transformers_version='4.21.1',
    pytorch_version='1.11.0',
    hyperparameters=hyperparameters,
    compiler_config=TrainingCompilerConfig(),
    disable_profiler=True,
    debugger_hook_config=False
)

pytorch_huggingface_estimator.fit()

Informationen zum Erstellen Ihres Trainingsskripts finden Sie auf den folgenden Seiten.

End-to-End-Beispiele finden Sie in den folgenden Notebooks:

Für PyTorch v1.12 können Sie verteiltes Training mit dem SageMaker Training Compiler ausführen, indem Sie die angegebene pytorch_xla-Option dem distribution-Parameter der PyTorch-Schätzerklasse von SageMaker AI hinzufügen.

from sagemaker.pytorch import PyTorch, TrainingCompilerConfig

# choose an instance type, specify the number of instances you want to use,
# and set the num_gpus variable the number of GPUs per instance.
instance_count=1
instance_type='ml.p3.8xlarge'
num_gpus=4

# the original max batch size that can fit to GPU memory without compiler
batch_size_native=16
learning_rate_native=float('5e-5')

# an updated max batch size that can fit to GPU memory with compiler
batch_size=26

# update learning rate
learning_rate=learning_rate_native/batch_size_native*batch_size*num_gpus*instance_count

hyperparameters={
    "n_gpus": num_gpus,
    "batch_size": batch_size,
    "learning_rate": learning_rate
}

pytorch_estimator=PyTorch(
    entry_point='your_training_script.py',
    source_dir='path-to-requirements-file', # Optional. Add this if need to install additional packages.
    instance_count=instance_count,
    instance_type=instance_type,
    framework_version='1.13.1',
    py_version='py3',
    hyperparameters=hyperparameters,
    compiler_config=TrainingCompilerConfig(),
    distribution ={'pytorchxla' : { 'enabled': True }},
    disable_profiler=True,
    debugger_hook_config=False
)

pytorch_estimator.fit()

Für PyTorch v1.11 und höher ist der SageMaker Training Compiler für verteiltes Training mit der im Parameter distribution angegebenen pytorch_xla Option verfügbar.

from sagemaker.huggingface import HuggingFace, TrainingCompilerConfig

# choose an instance type, specify the number of instances you want to use,
# and set the num_gpus variable the number of GPUs per instance.
instance_count=1
instance_type='ml.p3.8xlarge'
num_gpus=4

# the original max batch size that can fit to GPU memory without compiler
batch_size_native=16
learning_rate_native=float('5e-5')

# an updated max batch size that can fit to GPU memory with compiler
batch_size=26

# update learning rate
learning_rate=learning_rate_native/batch_size_native*batch_size*num_gpus*instance_count

hyperparameters={
    "n_gpus": num_gpus,
    "batch_size": batch_size,
    "learning_rate": learning_rate
}

pytorch_huggingface_estimator=HuggingFace(
    entry_point='your_training_script.py',
    instance_count=instance_count,
    instance_type=instance_type,
    transformers_version='4.21.1',
    pytorch_version='1.11.0',
    hyperparameters=hyperparameters,
    compiler_config=TrainingCompilerConfig(),
    distribution ={'pytorchxla' : { 'enabled': True }},
    disable_profiler=True,
    debugger_hook_config=False
)

pytorch_huggingface_estimator.fit()

Für die unterstützte Version von PyTorch v1.10.2 und früher benötigt der SageMaker Training Compiler einen alternativen Mechanismus zum Starten eines verteilten Trainingsjobs. Um verteiltes Training durchzuführen, verlangt der SageMaker Training Compiler, dass Sie ein SageMaker-AI-Startskript für verteilte Trainings an das entry_point-Argument übergeben und Ihr Trainingsskript an das hyperparameters-Argument übergeben. Das folgende Code-Beispiel veranschaulicht, wie Sie einen Hugging-Face-Schätzer von SageMaker AI konfigurieren, indem Sie die erforderlichen Änderungen anfügen.

from sagemaker.huggingface import HuggingFace, TrainingCompilerConfig

# choose an instance type, specify the number of instances you want to use,
# and set the num_gpus variable the number of GPUs per instance.
instance_count=1
instance_type='ml.p3.8xlarge'
num_gpus=4

# the original max batch size that can fit to GPU memory without compiler
batch_size_native=16
learning_rate_native=float('5e-5')

# an updated max batch size that can fit to GPU memory with compiler
batch_size=26

# update learning rate
learning_rate=learning_rate_native/batch_size_native*batch_size*num_gpus*instance_count

training_script="your_training_script.py"

hyperparameters={
    "n_gpus": num_gpus,
    "batch_size": batch_size,
    "learning_rate": learning_rate,
    "training_script": training_script     # Specify the file name of your training script.
}

pytorch_huggingface_estimator=HuggingFace(
    entry_point='distributed_training_launcher.py',    # Specify the distributed training launcher script.
    instance_count=instance_count,
    instance_type=instance_type,
    transformers_version='4.17.0',
    pytorch_version='1.10.2',
    hyperparameters=hyperparameters,
    compiler_config=TrainingCompilerConfig(),
    disable_profiler=True,
    debugger_hook_config=False
)

pytorch_huggingface_estimator.fit()

Das Startskript sollte wie folgt aussehen. Es umschließt Ihr Trainingsskript und konfiguriert die verteilte Trainingsumgebung in Abhängigkeit von der Größe der Trainings-Instance Ihrer Wahl.

# distributed_training_launcher.py

#!/bin/python

import subprocess
import sys

if __name__ == "__main__":
    arguments_command = " ".join([arg for arg in sys.argv[1:]])
    """
    The following line takes care of setting up an inter-node communication
    as well as managing intra-node workers for each GPU.
    """
    subprocess.check_call("python -m torch_xla.distributed.sm_dist " + arguments_command, shell=True)