Skip to content

Bitwise-perfect method for (de)serializing tensors in base64 #93859

@vadimkantorov

Description

@vadimkantorov

🚀 The feature, motivation and pitch

  • Convenient for pasting directly into repro code/comment on GitHub without attaching other files (also considering severe file extension limitations of GitHub file attachment)
  • Currently (at least, some time ago), some properties of Tensors (such as coalesced-ness) might be lost after existing torch.save+torch.load roundtrip, so need a bitwise perfect C++ Tensor object reconstruction with its layout/sparse structures/bitfields/storage/strides for reproducing/demonstrating problems with these, so a concise text-based format (typically for small tensors) would be useful!

Original context and proposal in:

Related on general use of utility for base64 torch.save/torch.load (?) format for repro purposes:

cc @mruberry @pearu

Alternatives

No response

Additional context

No response

Metadata

Metadata

Assignees

No one assigned

    Labels

    featureA request for a proper, new feature.module: serializationIssues related to serialization (e.g., via pickle, or otherwise) of PyTorch objectstriagedThis issue has been looked at a team member, and triaged and prioritized into an appropriate module

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions