PyTorch Serialized File Format

Format Description Categories >> Browse Alphabetical List

PyTorch Serialized File Format

Table of Contents

Identification and description
Local use
Sustainability factors
Quality and functionality factors
File type signifiers
Notes
Format specifications
Useful references

Format Description Properties

ID: fdd000644
Short name: PyTorch
Content categories: other
Format Category: file-format
Other facets: container-bundle, binary, structured
Last significant FDD update: 2025-04-02
Draft status: Preliminary

Identification and description

Relationship to other formats
Full name	PyTorch Serialized File Format
Description	The PyTorch Serialized File Format is an uncompressed ZIP64 archive (as of PyTorch version 1.6.0) that is used to save the weights and biases of a trained PyTorch model to disk, along with other data. It is common convention to save PyTorch models using either .pt or .pth as the file extension, but not mandatory. The file within the archive that contains the model's weights and biases will have the file extension .pkl, having been pickled/serialized upon save. PyTorch is an open-source, Python-based deep learning library based on the Torch library. According to Datacamp's Deep Learning with PyTorch Cheat Sheet, PyTorch models are used for neural network development, natural language processing, computer vision, and reinforcement learning. The PyTorch Serialized File Format is documented within the PyTorch documentation. A PyTorch model is saved to disk using the torch.save function. This function uses Python's pickle utility to serialize the object that is passed to it. When saving a PyTorch model for inference, PyTorch recommends only saving the trained model's learned parameters (i.e., its weights and biases), which are found in its state_dict. According to PyTorch, a state_dict is "a Python dictionary object that maps each layer to its parameter tensor. Note that only layers with learnable parameters (convolutional layers, linear layers, etc.) and registered buffers (batchnorm's running_mean) have entries in the model's state_dict." Inside the archive, as described in PyTorch's serialization semantics documentation, is: A pickled/serialized version of the object that was passed to torch.save, such as a state_dict (excluding its storage objects, which are stored in the data directory) A string with the sys.byteorder ("little" or "big") A data directory containing a file for each storage in the object Version number at save time Comments welcome.
Production phase	May be used at any lifecycle phase for saving PyTorch models.
Contains	Serialized Python object, not described separately at this time.

Local use

LC experience or existing holdings	As of this writing in April 2025, the Library of Congress does not have PyTorch serialized files in its collections.
LC preference	The Library of Congress has not yet expressed any format preference for preserving machine learning models.

Sustainability factors

Disclosure	PyTorch is an open-source library with documentation available on the PyTorch website and code available on the PyTorch GitHub repository.
Documentation	PyTorch documents its serialized file format for torch.save on the serialization semantics page.
Adoption	The PyTorch blog described PyTorch in 2022 as "one of the primary platforms for AI research, as well as commercial production use." Also in 2022, The Register called PyTorch "one of the major deep learning frameworks at the moment, the other being TensorFlow, developed by Google." A 2017 article from O'Reilly Media credits the adoption of PyTorch "to native Python-style imperative programming already familiar to researchers, data scientists, and developers of popular Python libraries such as NumPy and SciPy." Comments welcome.
Licensing and patents	From the PyTorch GitHub repository: "PyTorch has a BSD-style license, as found in the LICENSE file."
Transparency	The PyTorch Serialized File Format is not intended to be manually opened, but to be imported into a PyTorch module according to FileInfo.com. See ZIP_PK for more general information. Comments welcome.
Self-documentation	PyTorch ZIP archives contain some metadata: a string with the sys.byteorder ("little" or "big") and a version number at save time that can be used at load time. See ZIP_PK for more general information. Comments welcome. Accessibility Features The PyTorch Serialized File Format has no specific attributes to support accessibility. Comments welcome.
External dependencies	PyTorch can be installed on Windows, Mac, or Linux OS. Prerequisites for installation are outlined on the PyTorch Get Started page. The PyTorch documentation on saving and loading models recommends saving only the model's learnable parameters, which are stored in the state_dict object of the model's module, over saving complete model architecture and parameters. Saving the entire model is not recommended because it creates dependencies on the class definitions and directory structure used save time. Comments welcome.
Technical protection considerations	TBD

Quality and functionality factors

Aggregate
Compression	See ZIP_PK.
Support for Error Dectection	See ZIP_PK.
Beyond normal functionality	See ZIP_PK.

File type signifiers and format identifiers

Tag	Value	Note
Filename extension	pt pth	From PyTorch's Saving and Loading Models: Use of these extensions is common convention when saving PyTorch models.
Other	See note.	NARA File Format Preservation Plan ID has no corresponding entry as of April 2025.
Pronom PUID	See note.	PRONOM has no corresponding entry as of April 2025.
Wikidata Title ID	Q47509047	See https://www.wikidata.org/wiki/Q47509047 for PyTorch library.

Notes

General	The two defining features of PyTorch, according to the PyTorch README on GitHub, are "tensor computation with strong GPU acceleration" and "deep neural networks built on a tape-based autograd system." From the PyTorch tutorial on tensors: "Tensors are a specialized data structure that are very similar to arrays and matrices. In PyTorch, we use tensors to encode the inputs and outputs of a model, as well as the model's parameters." Tensors are passed through a neural network, which the PyTorch tutorial on building models describes as a module that contains other modules nested within it. These layers of modules perform operations on the tensor data. Modules can have parameters, "weights and biases that are optimized during training," which are also encoded as tensors.
History	PyTorch was originally developed by the Facebook AI Research (FAIR) team at Meta and first launched in 2017 according to the FAIR blog. Since September 2022, it has been governed, developed, and maintained by the PyTorch Foundation, part of the Linux Foundation, per an announcement on the PyTorch blog. As of this writing in April 2025, PyTorch's latest stable release is version 2.6.

General

The two defining features of PyTorch, according to the PyTorch README on GitHub, are "tensor computation with strong GPU acceleration" and "deep neural networks built on a tape-based autograd system."

From the PyTorch tutorial on tensors: "Tensors are a specialized data structure that are very similar to arrays and matrices. In PyTorch, we use tensors to encode the inputs and outputs of a model, as well as the model's parameters."

Tensors are passed through a neural network, which the PyTorch tutorial on building models describes as a module that contains other modules nested within it. These layers of modules perform operations on the tensor data. Modules can have parameters, "weights and biases that are optimized during training," which are also encoded as tensors.

History

PyTorch was originally developed by the Facebook AI Research (FAIR) team at Meta and first launched in 2017 according to the FAIR blog. Since September 2022, it has been governed, developed, and maintained by the PyTorch Foundation, part of the Linux Foundation, per an announcement on the PyTorch blog. As of this writing in April 2025, PyTorch's latest stable release is version 2.6.

Format specifications

PyTorch documentation (https://pytorch.org/docs/stable/index.html).
Serialized file format for torch.save (https://pytorch.org/docs/stable/notes/serialization.html#serialized-file-format-for-torch-save).

Useful references

URLs

PyTorch
- PyTorch Foundation (https://pytorch.org/).
- Build the Neural Network, PyTorch tutorials (https://pytorch.org/tutorials/beginner/basics/buildmodel_tutorial.html).
- A Gentle Introduction to torch.autograd, PyTorch tutorials (https://pytorch.org/tutorials/beginner/blitz/autograd_tutorial.html).
- Get Started (https://pytorch.org/get-started/locally/).
- pytorch GitHub (https://github.com/pytorch/pytorch).
- pytorch/LICENSE, GitHub (https://github.com/pytorch/pytorch/blob/main/LICENSE).
- PyTorch strengthens its governance by joining the Linux Foundation, PyTorch blog (https://pytorch.org/blog/PyTorchfoundation/).
- Saving and Loading Models, PyTorch tutorials (https://pytorch.org/tutorials/beginner/saving_loading_models.html).
- Serialization semantics, PyTorch documentation (https://pytorch.org/docs/stable/notes/serialization.html).
- Tensors, PyTorch tutorials (https://pytorch.org/tutorials/beginner/basics/tensorqs_tutorial.html).
- Unlocking the Latest Features in PyTorch 2.6 for Intel Platforms, PyTorch blog (https://pytorch.org/blog/unlocking-pt-2-6-intel/).
All about PKL Files, FILExt (https://filext.com/file-extension/PKL).
Comprehending .pth Files: The Backbone of PyTorch Models, Medium (https://medium.com/@manasnandmohan/comprehending-pth-files-the-backbone-of-pytorch-models-ef9b232e092a).
Deep Learning with PyTorch Cheat Sheet, DataCamp (https://www.datacamp.com/cheat-sheet/deep-learning-with-py-torch).
Facebook brings GPU-powered machine learning to Python, InfoWorld, (https://web.archive.org/web/20170119223337/https:/www.infoworld.com/article/3159120/artificial-intelligence/facebook-brings-gpu-powered-machine-learning-to-python.html). Available via Internet Archive.
FAIR turns five: What we've accomplished and where we're headed, Meta (https://ai.meta.com/blog/fair-turns-five-what-weve-accomplished-and-where-were-headed/).
How to save and load models in PyTorch, Weights %26 Biases (https://wandb.ai/wandb/common-ml-errors/reports/How-to-save-and-load-models-in-PyTorch--VmlldzozMjg0MTE).
List of ML File Formats (https://github.com/trailofbits/ml-file-formats).
NLP with PyTorch: A Comprehensive Guide, DataCamp (https://www.datacamp.com/tutorial/nlp-with-pytorch-a-comprehensive-guide).
Opening PTH Files, FILExt (https://filext.com/file-extension/PTH).
pickle — Python object serialization, Python (https://docs.python.org/3/library/pickle.html).
.PT File Extension, FileInfo.com (https://fileinfo.com/extension/pt).
.PTH File Extension, FileInfo.com (https://fileinfo.com/extension/pth).
PyTorch gets lit under The Linux Foundation, The Register (https://www.theregister.com/2022/09/12/pytorch_meta_linux_foundation/).
Torch (http://torch.ch/).
Understanding File Extensions in PyTorch: .pt, .pth, and .pwf, GeeksforGeeks (https://www.geeksforgeeks.org/understanding-file-extensions-in-pytorch-pt-pth-and-pwf/).
What exactly is the .pth file? Medium (https://medium.com/@yulin_li/what-exactly-is-the-pth-file-9a487044a36b).
When two trends fuse: PyTorch and recommender systems, O'Reilly Media (https://www.oreilly.com/content/when-two-trends-fuse-pytorch-and-recommender-systems/).
Why AI and machine learning researchers are beginning to embrace PyTorch, O'Reilly Media (https://www.oreilly.com/radar/podcast/why-ai-and-machine-learning-researchers-are-beginning-to-embrace-pytorch/).
Wikipedia articles
- PyTorch (https://en.wikipedia.org/wiki/PyTorch).
- BSD licenses (https://en.wikipedia.org/wiki/BSD_licenses).
- Computer vision (https://en.wikipedia.org/wiki/Computer_vision).
- Deep learning (https://en.wikipedia.org/wiki/Deep_learning).
- Large language model (https://en.wikipedia.org/wiki/Large_language_model).
- Natural language processing (https://en.wikipedia.org/wiki/Natural_language_processing).
- Neural network (machine learning) (https://en.wikipedia.org/wiki/Neural_network_(machine_learning)).
- Reinforcement learning (https://en.wikipedia.org/wiki/Reinforcement_learning).
- Serialization (https://en.wikipedia.org/wiki/Serialization).
- Tensor (machine learning) (https://en.wikipedia.org/wiki/Tensor_(machine_learning)).
- Torch (machine learning) (https://en.wikipedia.org/wiki/Torch_(machine_learning)).
- ZIP (file format) (https://en.wikipedia.org/wiki/ZIP_(file_format)).
Wikidata entry for Q47509047 (https://www.wikidata.org/wiki/Q47509047). Information in Wikidata about PyTorch. Wikidata Title ID: Q281876.

Last Updated: 04/02/2025

Sustainability of Digital Formats: Planning for Library of Congress Collections

Introduction \| Sustainability Factors \| Content Categories \| Format Descriptions \| Contact

Sustainability of Digital Formats: Planning for Library of Congress Collections