n_pth.h File Reference

Reading the tensors of a PyTorch checkpoint (.pth, .pt) without Python. More...

#include <stddef.h>
#include <stdint.h>

Include dependency graph for n_pth.h:

This graph shows which files directly or indirectly include this file:

Go to the source code of this file.

Data Structures
struct	N_PTH
	A checkpoint's tensors. More...
struct	N_PTH_TENSOR
	One tensor of a checkpoint. More...

Macros
#define	N_PTH_MAX_DIMS   8
	Most dimensions a tensor read may have.

Functions
const N_PTH_TENSOR *	n_pth_find (const N_PTH *pth, const char *name)
	find a tensor by its path
void	n_pth_free (N_PTH **pth)
	release a checkpoint's tensors
N_PTH *	n_pth_load (const char *path)
	read every tensor of a checkpoint

Detailed Description

Reading the tensors of a PyTorch checkpoint (.pth, .pt) without Python.

A checkpoint saved by torch.save() since PyTorch 1.6 is a ZIP archive: a pickle, data.pkl, that describes the saved object, and one file a storage holding the raw numbers. This reads the pickle with a small interpreter that builds data only: dictionaries, lists, tuples, numbers, strings, and the three constructors a state dictionary uses (collections.OrderedDict, torch._utils._rebuild_tensor_v2 and _rebuild_parameter). Nothing named in the file is ever called; any other constructor becomes an opaque value, so a checkpoint's extra metadata does not stop its tensors being read.

Every tensor reachable through dictionaries is returned by its path, the keys joined by dots ("state_dict.backbone.0.0.weight"), as floats: float32, float16, bfloat16 and float64 storages are read and converted. Strided views are gathered. Little endian storages only, which is what every platform PyTorch runs on writes.

Every length, index and count in the file is checked against the data actually there: a damaged or hostile file is refused, not trusted.

Author

Castagnier Mickael

Version

1.0

Date

29/09/2026

Definition in file n_pth.h.