Amazon S3 Connector for PyTorch

The Amazon S3 Connector for PyTorch delivers high throughput for PyTorch training jobs that access or store data in Amazon S3. Using the S3 Connector for PyTorch automatically optimizes performance when downloading training data from and writing checkpoints to Amazon S3, eliminating the need to write your own code to list S3 buckets and manage concurrent requests.

Amazon S3 Connector for PyTorch provides implementations of PyTorch's dataset primitives that you can use to load training data from Amazon S3. It supports both map-style datasets for random data access patterns and iterable-style datasets for streaming sequential data access patterns. The S3 Connector for PyTorch also includes a checkpointing interface to save and load checkpoints directly to Amazon S3, without first saving to local storage.

Getting Started

Prerequisites

• Python 3.8-3.14 is supported.
  • Note: Python 3.8 support will be deprecated in a future release, see #399
• PyTorch >= 2.0 (TODO: Check with PyTorch 1.x)

Installation

pip install s3torchconnector

Amazon S3 Connector for PyTorch supports pre-built wheels via Pip only for Linux and MacOS for now. (Note: macOS x86_64 wheel support will be deprecated in a future release, see #398) For other platforms, see DEVELOPMENT for build instructions.

Configuration

To use s3torchconnector, AWS credentials must be provided through one of the following methods:

• EC2 Instance Role: If you are using this library on an EC2 instance, specify an IAM role and then give the EC2 instance access to that role.
• AWS CLI: Install and configure awscli and run aws configure.
• AWS Credential Files: Set credentials in the AWS credentials profile file on the local system, located at: ~/.aws/credentials on Unix or macOS.
• Environment Variables: Set the AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY environment variables.

To use a specific AWS profile configured in ~/.aws/config and ~/.aws/credentials, you can either:

• Set environment variable AWS_PROFILE=custom-profile, or
• Pass the profile name to the S3ClientConfig object, e.g. S3ClientConfig(profile="custom-profile").

For a more detailed configuration guide, see AWS CLI docs.

Examples

API docs are showing API of the public components. End to end example of how to use s3torchconnector can be found under the examples directory.

Sample Examples

The simplest way to use the S3 Connector for PyTorch is to construct a dataset, either a map-style or iterable-style dataset, by specifying an S3 URI (a bucket and optional prefix) and the region the bucket is located in:

from s3torchconnector import S3MapDataset, S3IterableDataset

# You need to update <BUCKET> and <PREFIX>
DATASET_URI="s3://<BUCKET>/<PREFIX>"
REGION = "us-east-1"

iterable_dataset = S3IterableDataset.from_prefix(DATASET_URI, region=REGION)

# Datasets are also iterators. 
for item in iterable_dataset:
  print(item.key)

# S3MapDataset eagerly lists all the objects under the given prefix 
# to provide support of random access.  
# S3MapDataset builds a list of all objects at the first access to its elements or 
# at the first call to get the number of elements, whichever happens first.
# This process might take some time and may give the impression of being unresponsive.
map_dataset = S3MapDataset.from_prefix(DATASET_URI, region=REGION)

# Randomly access an item in map_dataset.
item = map_dataset[0]

# Learn about bucket, key, and content of the object
bucket = item.bucket
key = item.key
content = item.read()
len(content)

In addition to data loading primitives, the S3 Connector for PyTorch also provides an interface for saving and loading model checkpoints directly to and from an S3 bucket.

from s3torchconnector import S3Checkpoint

import torchvision
import torch

CHECKPOINT_URI="s3://<BUCKET>/<KEY>/"
REGION = "us-east-1"
checkpoint = S3Checkpoint(region=REGION)

model = torchvision.models.resnet18()

# Save checkpoint to S3
with checkpoint.writer(CHECKPOINT_URI + "epoch0.ckpt") as writer:
    torch.save(model.state_dict(), writer)

# Load checkpoint from S3
with checkpoint.reader(CHECKPOINT_URI + "epoch0.ckpt") as reader:
    state_dict = torch.load(reader)

model.load_state_dict(state_dict)

Using datasets or checkpoints with Amazon S3 Express One Zone directory buckets requires only to update the URI, following base-name--azid--x-s3 bucket name format. For example, assuming the following directory bucket name my-test-bucket--usw2-az1--x-s3 with the Availability Zone ID usw2-az1, then the URI used will look like: s3://my-test-bucket--usw2-az1--x-s3/<PREFIX> (please note that the prefix for Amazon S3 Express One Zone should end with '/'), paired with region us-west-2.

Distributed checkpoints

Overview

Amazon S3 Connector for PyTorch provides robust support for PyTorch distributed checkpoints. This feature includes:

• S3StorageWriter: Implementation of PyTorch's StorageWriter interface.

• S3StorageReader: Implementation of PyTorch's StorageReader interface.

  • Supports configurable reading strategies via the reader_constructor parameter (see Reader Configurations).
  • Uses DCPOptimizedS3Reader by default for faster loading and partial checkpoint optimizations.
  • Please refer to DCPOptimizedS3Reader Errors for troubleshooting.

• S3FileSystem: An implementation of PyTorch's FileSystemBase.

These tools enable seamless integration of Amazon S3 with PyTorch Distributed Checkpoints, allowing efficient storage and retrieval of distributed model checkpoints.

Prerequisites and Installation

PyTorch 2.3 or newer is required. To use the distributed checkpoints feature, install S3 Connector for PyTorch with the dcp extra:

pip install s3torchconnector[dcp]

Sample Example

End-to-end examples for using distributed checkpoints with S3 Connector for PyTorch can be found in the examples/dcp directory.

from s3torchconnector.dcp import S3StorageWriter, S3StorageReader

import torchvision
import torch.distributed.checkpoint as DCP

# Configuration
CHECKPOINT_URI = "s3://<BUCKET>/<KEY>/"
REGION = "us-east-1"

model = torchvision.models.resnet18()

# Save distributed checkpoint to S3
s3_storage_writer = S3StorageWriter(
    region=REGION, 
    path=CHECKPOINT_URI,
    thread_count=8, # optional; number of IO threads to use to write
) 
DCP.save(
    state_dict=model.state_dict(),
    storage_writer=s3_storage_writer,
)

# Load distributed checkpoint from S3
# S3StorageReader uses DCPOptimizedS3Reader by default for improved performance
model = torchvision.models.resnet18()
model_state_dict = model.state_dict()
s3_storage_reader = S3StorageReader(
    region=REGION, 
    path=CHECKPOINT_URI,
)
DCP.load(
    state_dict=model_state_dict,
    storage_reader=s3_storage_reader,
)
model.load_state_dict(model_state_dict)

S3 Prefix Strategies for Distributed Checkpointing

S3StorageWriter implements various prefix strategies to optimize checkpoint organization in S3 buckets. These strategies are specifically designed to prevent throttling (503 Slow Down errors) in high-throughput scenarios by implementing S3 key naming best practices as outlined in Best practices design patterns: optimizing Amazon S3 performance.

When many distributed training processes write checkpoints simultaneously, the prefixing strategies help distribute the load across multiple S3 partitions.

Available Strategies

1. RoundRobinPrefixStrategy

Distributes checkpoints across specified prefixes in a round-robin fashion, ideal for balancing data across multiple storage locations.

from s3torchconnector.dcp import RoundRobinPrefixStrategy, S3StorageWriter

model = torchvision.models.resnet18()

# Initialize with multiple prefixes and optional epoch tracking
strategy = RoundRobinPrefixStrategy(
    user_prefixes=["shard1", "shard2", "shard3"],
    epoch_num=5  # Optional: for checkpoint versioning
)

writer = S3StorageWriter(
    region=REGION,
    path="CHECKPOINT_URI",
    prefix_strategy=strategy
)

# Save checkpoint
DCP.save(
    state_dict=model.state_dict(),
    storage_writer=writer
)

Output Structure:

CHECKPOINT_URI
├── shard1/
│   └── epoch_5/
│       ├── __0_0.distcp
│       ├── __3_0.distcp
│       └── ...
├── shard2/
│   └── epoch_5/
│       ├── __1_0.distcp
│       ├── __4_0.distcp
│       └── ...
└── shard3/
    └── epoch_5/
        ├── __2_0.distcp
        ├── __5_0.distcp
        └── ...

2. BinaryPrefixStrategy

Generates binary (base-2) prefixes for optimal partitioning in distributed environments.

from s3torchconnector.dcp import BinaryPrefixStrategy

strategy = BinaryPrefixStrategy(
    epoch_num=1,          # Optional: for checkpoint versioning
    min_prefix_len=10     # Optional: minimum prefix length
)

Output Structure:

s3://my-bucket/checkpoints/
├── 0000000000/
│   └── epoch_1/
│       └── __0_0.distcp
├── 1000000000/
│   └── epoch_1/
│       └── __1_0.distcp
├── 0100000000/
│   └── epoch_1/
│       └── __2_0.distcp
└── ...

3. HexPrefixStrategy

Uses hexadecimal (base-16) prefixes for a balance of efficiency and readability.

from s3torchconnector.dcp import HexPrefixStrategy

strategy = HexPrefixStrategy(
    epoch_num=1,          # Optional: for checkpoint versioning
    min_prefix_len=4      # Optional: minimum prefix length
)

Output Structure:

s3://my-bucket/checkpoints/
├── 0000/
│   └── epoch_1/
│       └── __0_0.distcp
├── 1000/
│   └── epoch_1/
│       └── __1_0.distcp
...
├── f000/
│   └── epoch_1/
│       └── __15_0.distcp
└── ...

Creating Custom Strategies

You can implement custom prefix strategies by extending the S3PrefixStrategyBase class:

from s3torchconnector.dcp import S3PrefixStrategyBase

class CustomPrefixStrategy(S3PrefixStrategyBase):
    def __init__(self, custom_param):
        super().__init__()
        self.custom_param = custom_param

    def generate_prefix(self, rank: int) -> str:
        return f"custom_{self.custom_param}/{rank}/"

Parallel/Distributed Training

Amazon S3 Connector for PyTorch provides support for parallel and distributed training with PyTorch, allowing you to leverage multiple processes and nodes for efficient data loading and training. Both S3IterableDataset and S3MapDataset can be used for this purpose.

S3IterableDataset

The S3IterableDataset can be directly passed to PyTorch's DataLoader for parallel and distributed training. By default, all worker processes will share the same list of training objects. However, if you need each worker to have access to a unique portion of the dataset for better parallelization, you can enable dataset sharding using the enable_sharding parameter.

dataset = S3IterableDataset.from_prefix(DATASET_URI, region=REGION, enable_sharding=True)
dataloader = DataLoader(dataset, num_workers=4)

When enable_sharding is set to True, the dataset will be automatically sharded across available number of workers. This sharding mechanism supports both parallel training on a single host and distributed training across multiple hosts. Each worker, regardless of its host, will load and process a distinct subset of the dataset.

S3MapDataset

For the S3MapDataset, you need to pass it to DataLoader along with a DistributedSampler wrapped around it. The DistributedSampler ensures that each worker or node receives a unique subset of the dataset, enabling efficient parallel and distributed training.

dataset = S3MapDataset.from_prefix(DATASET_URI, region=REGION)
sampler = DistributedSampler(dataset)
dataloader = DataLoader(dataset, sampler=sampler, num_workers=4)

Lightning Integration

Amazon S3 Connector for PyTorch includes an integration for PyTorch Lightning, featuring S3LightningCheckpoint, an implementation of Lightning's CheckpointIO. This allows users to make use of Amazon S3 Connector for PyTorch's S3 checkpointing functionality with PyTorch Lightning.

Getting Started

Installation

pip install s3torchconnector[lightning]

Examples

End to end examples for the PyTorch Lightning integration can be found in the examples/lightning directory.

from lightning import Trainer
from s3torchconnector.lightning import S3LightningCheckpoint

...

s3_checkpoint_io = S3LightningCheckpoint("us-east-1")
trainer = Trainer(
    plugins=[s3_checkpoint_io],
    default_root_dir="s3://bucket_name/key_prefix/"
)
trainer.fit(model)

Using S3 Versioning to Manage Checkpoints

When working with model checkpoints, you can use the S3 Versioning feature to preserve, retrieve, and restore every version of your checkpoint objects. With versioning, you can recover more easily from unintended overwrites or deletions of existing checkpoint files due to incorrect configuration or multiple hosts accessing the same storage path.

When versioning is enabled on an S3 bucket, deletions insert a delete marker instead of removing the object permanently. The delete marker becomes the current object version. If you overwrite an object, it results in a new object version in the bucket. You can always restore the previous version. See Deleting object versions from a versioning-enabled bucket for more details on managing object versions.

To enable versioning on an S3 bucket, see Enabling versioning on buckets. Normal Amazon S3 rates apply for every version of an object stored and transferred. To customize your data retention approach and control storage costs for earlier versions of objects, use object versioning with S3 Lifecycle.

S3 Versioning and S3 Lifecycle are not supported by S3 Express One Zone.

Direct S3Client Usage

For advanced use cases, you can use the S3Client directly for custom streaming patterns and integration with existing pipelines.

from s3torchconnector._s3client import S3Client

REGION = "us-east-1"
BUCKET_NAME = "my-bucket"
OBJECT_KEY = "large_object.bin"

s3_client = S3Client(region=REGION)

# Writing data to S3
data = b"content" * 1048576
s3writer = s3_client.put_object(bucket=BUCKET_NAME, key=OBJECT_KEY)
s3writer.write(data)
s3writer.close()

# Reading data from S3
s3reader = s3_client.get_object(bucket=BUCKET_NAME, key=OBJECT_KEY)
data = s3reader.read()

Reader Configurations

Amazon S3 Connector for PyTorch supports three types of readers, configurable through S3ReaderConstructor.

Reader Types

1. Sequential Reader

• Default for non-DCP use cases.
• Downloads and buffers the entire S3 object in memory.
• Prioritizes performance over memory usage by buffering entire objects.

2. Range-based Reader

• Performs byte-range requests to read specific portions of S3 objects without downloading the entire object.
• Prioritizes memory efficiency, with performance gains only for sparse partial reads in large objects.
• Features adaptive buffering with forward overlap handling:
  • Small reads (< buffer_size): Use internal buffer to reduce S3 API calls.
  • Large reads (≥ buffer_size): Bypass buffer for direct transfer.

3. DCP-Optimized Reader

• Default for PyTorch Distributed Checkpoint (DCP) loading with S3StorageReader.
• Provides performance improvements through per-item buffers and zero-copy buffer management.
• Enables efficient partial checkpoint loading (e.g. model-only) through selective data fetching with range coalescing.
• Automatically handles range metadata injection from DCP load plan.
• Requires sequential access patterns (automatically enforced in S3StorageReader.prepare_local_plan())

When to Use Each Reader

• Sequential Reader: For processing entire objects, and when repeated access to the data is required. Best for most general use cases.
• Range-based Reader: For larger objects (100MB+) that require sparse partial reads, and in memory-constrained environments.
• DCP-Optimized Reader: For typical PyTorch Distributed Checkpoint loading scenarios for highest performance and memory-efficiency. (Default for S3StorageReader)

Note: S3Reader instances are not thread-safe and should not be shared across threads. For multiprocessing with DataLoader, each worker process creates its own S3Reader instance automatically.

Examples

For S3ReaderConstructor usage details, please refer to the S3ReaderConstructor documentation. Below are some examples for S3ReaderConstructor usage.

Direct method - S3Client usage with range-based reader without buffer:

# Direct S3Client usage for zero-copy partial reads into pre-allocated buffers, for memory efficiency and fast data transfer
from s3torchconnector._s3client import S3Client
from s3torchconnector import S3ReaderConstructor

s3_client = S3Client(region=REGION)
reader_constructor = S3ReaderConstructor.range_based(
    buffer_size=0  # No buffer, for direct transfer
)
s3reader = s3_client.get_object(
    bucket=BUCKET_NAME, 
    key=OBJECT_NAME, 
    reader_constructor=reader_constructor
)

buffer = bytearray(10 * 1024 * 1024)  # 10MB buffer
s3reader.seek(100 * 1024 * 1024)   # Skip to 100MB offset
bytes_read = s3reader.readinto(buffer)  # Direct read into buffer

DCP interface - S3StorageReader usage with dcp-optimized reader:

# Load checkpoint with dcp-optimized reader for better performance
from s3torchconnector.dcp import S3StorageReader
from s3torchconnector import S3ReaderConstructor

# dcp_optimized is already the default for S3StorageReader; demonstration purposes only. 
reader_constructor = S3ReaderConstructor.dcp_optimized()
s3_storage_reader = S3StorageReader(
    region=REGION, 
    path=CHECKPOINT_URI,
    reader_constructor=reader_constructor
)
DCP.load(
    state_dict=model_state_dict,
    storage_reader=s3_storage_reader,
)

Dataset interface - S3MapDataset usage with sequential reader:

# Use sequential reader for optimal performance when reading entire objects
from s3torchconnector import S3MapDataset, S3ReaderConstructor

dataset = S3MapDataset.from_prefix(
    DATASET_URI, 
    region=REGION,
    reader_constructor=S3ReaderConstructor.sequential()
)

for item in dataset:
    content = item.read()
    ...

Contributing

We welcome contributions to Amazon S3 Connector for PyTorch. Please see CONTRIBUTING for more information on how to report bugs or submit pull requests.

Development

See DEVELOPMENT for information about code style, development process, and guidelines.

Compatibility with other storage services

S3 Connector for PyTorch delivers high throughput for PyTorch training jobs that access or store data in Amazon S3. While it may be functional against other storage services that use S3-like APIs, they may inadvertently break when we make changes to better support Amazon S3. We welcome contributions of minor compatibility fixes or performance improvements for these services if the changes can be tested against Amazon S3.

Security issue notifications

If you discover a potential security issue in this project we ask that you notify AWS Security via our vulnerability reporting page.

Code of conduct

This project has adopted the Amazon Open Source Code of Conduct. See CODE_OF_CONDUCT.md for more details.

License

Amazon S3 Connector for PyTorch has a BSD 3-Clause License, as found in the LICENSE file.