.PT

Introduction

A .pt or .pth file is whatever PyTorch wrote with its torch.save function. Most often it is a model’s state_dict, a dictionary from layer names to weight tensors, but it can be any Python object that contains tensors: a whole model, an optimiser’s state or a training checkpoint. The two extensions mean the same thing. Hugging Face models from before safetensors are the same format named pytorch_model.bin.

Loading one uses Python’s pickle module, which can run code hidden in the file, so a model from an untrusted source is a risk. Since version 2.6, torch.load loads only tensors and plain data unless told otherwise. Shared models have largely moved to safetensors, which cannot run code.

Python itself also uses .pth for unrelated path configuration files in its site-packages folder.

Structure

Since PyTorch 1.6 the file is a ZIP archive whose entries are stored without compression, so tensors can be read in place. Everything sits in one top folder:

• data.pkl: the saved object, pickled, with each tensor replaced by a reference to a storage record
• data/0, data/1 and so on: the raw bytes of each storage, one entry per storage
• version: the version of the serialisation format
• byteorder in newer files: little or big

Older files, from before 1.6, are a series of pickles one after another followed by the storage bytes, with no ZIP wrapper. TorchScript models saved with torch.jit.save are ZIP archives too, with the model’s code in a code/ folder beside the data.

Also used for

• GGML model file .bin

Links