This change groups encoding and decoding logic for header and metadata entries into dedicated classes, completely separate from the TensorSerializer and TensorDeserializer classes. TensorSerializer was already mostly updated to do this in #46, so this mainly affects TensorDeserializer.
This disentangles most of the parsing of tensorizer's custom binary format from the use of that data, making it easier to update and optimize the deserializer's major bottlenecks like I/O, hash computation, and device transfer as was done in the serializer.
Deserialization Reorganization
This change groups encoding and decoding logic for header and metadata entries into dedicated classes, completely separate from the
TensorSerializer
andTensorDeserializer
classes.TensorSerializer
was already mostly updated to do this in #46, so this mainly affectsTensorDeserializer
.This disentangles most of the parsing of
tensorizer
's custom binary format from the use of that data, making it easier to update and optimize the deserializer's major bottlenecks like I/O, hash computation, and device transfer as was done in the serializer.