← All notes

Answer: Instead of saving a module directly, for compatibility reasons it is recommended to instead save only its state dict. (S1)

What gets saved

• Custom data structures that include PyTorch tensors can also be saved if the data structure is pickle-able. (S1)
• When PyTorch saves tensors it saves their storage objects and tensor metadata separately. (S1)
• When saving tensors with fewer elements than their storage objects, the size of the saved file can be reduced by first cloning the tensors. (S1)
• The mmap argument in torch.load() allows for lazy loading of tensor storages. (S1)

File format

• A common PyTorch convention is to save tensors using .pt file extension. (S2)
• The 1.6 release of PyTorch switched torch.save to use a new zipfile-based file format. (S2)
• torch.load still retains the ability to load files in the old format. (S2)

Claims

• id: C1 claim: Save the state dict, not the module. source: S1 evidence: "to instead save only its state dict."
• id: C2 claim: Storage and metadata are saved separately. source: S1 evidence: "When PyTorch saves tensors it saves their storage objects and tensor"
• id: C3 claim: Clone small views to shrink files. source: S1 evidence: "the saved file can be reduced by first cloning the tensors."
• id: C4 claim: torch.save uses a zipfile format since 1.6. source: S2 evidence: "The 1.6 release of PyTorch switched torch.save to use a new"

Sources

• [S1] Serialization semantics — PyTorch 2.15 documentation. PyTorch docs. https://docs.pytorch.org/docs/2.15/notes/serialization.html . License: BSD-style (pytorch/pytorch LICENSE; Copyright (c) 2016-present, Facebook Inc.; https://github.com/pytorch/pytorch/blob/main/LICENSE); docs only, not the tutorials repo; PyTorch name and logo not licensed. Accessed 2026-11-04.
• [S2] torch.save — PyTorch 2.15 documentation. PyTorch docs. https://docs.pytorch.org/docs/2.15/generated/torch.save.html . License: BSD-style (pytorch/pytorch LICENSE; Copyright (c) 2016-present, Facebook Inc.; https://github.com/pytorch/pytorch/blob/main/LICENSE); docs only, not the tutorials repo; PyTorch name and logo not licensed. Accessed 2026-11-04.

License: this note is CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Summary by iinventedlight (AI-assisted).

@misc{iil_note_24vtnypb,
  author = {{iinventedlight}},
  title = {{Saving and loading in PyTorch}},
  howpublished = {iinventedlight Library},
  year = {2026},
  date = {2026-11-04},
  url = {https://iinventedlight.com/library/notes/pytorch-serialization/},
  note = {Permanent ID iil:note:24vtnypb. License: CC-BY-4.0.},
}

Markdown

Spotted a mistake? Use the MCP feedback tool or email hello@iinventedlight.com.

For language models and agents: start at /llms.txt. This page in Markdown: /library/notes/pytorch-serialization.md.