latest (dev)
Copy
Latest development documentation · Updated 2026-10-08
tensorplay.serialization
Classes
Functions
Convert a model directory or a supported checkpoint into MEGA. |
|
Convert a model directory or a supported checkpoint into MEGA. |
|
Read MEGA metadata and tensor ranges without materializing tensors. |
|
Resolve one saved storage location. |
|
Format Constants
The serializer identifies its files with a magic number and a protocol
version in the header. Payloads are pickled with
tensorplay.serialization.DEFAULT_PROTOCOL unless save(..., pickle_protocol=...)
says otherwise, and storages inside mega archives are aligned to
tensorplay.serialization.DEFAULT_ALIGNMENT bytes unless save(..., alignment=...)
says otherwise. A mega archive is a directory whose name ends with
tensorplay.serialization.MEGA_EXTENSION, holding an index file ending with
tensorplay.serialization.MEGA_INDEX_SUFFIX next to the serialized data.
tensorplay.serialization.MAGIC_NUMBER— the magic value every archive starts with; loaders reject files whose header does not match it.tensorplay.serialization.PROTOCOL_VERSION— the header protocol version.tensorplay.serialization.DEFAULT_PROTOCOL— default pickle protocol for the payload.tensorplay.serialization.DEFAULT_ALIGNMENT— default byte alignment for storages written into mega archives.tensorplay.serialization.MEGA_EXTENSION— file suffix of a mega archive.tensorplay.serialization.MEGA_INDEX_SUFFIX— file suffix of a mega archive’s index document.
Help improve this page
Found an error, an unclear step, or a missing example?

