Skip to content

Support for multi-file ckpt #13

@Jackmin801

Description

@Jackmin801

Great work! Love the goals of the library and the blogpost was very insightful. Hoping to be using flashpack for the foreseeable future for all weight loading needs.

I was wondering if the lib currently supports or plans to support loading from multiple files similar to safetensors / dcp.
If its in the works, i was also wondering what the plan was to optimize for training and inference, particularly how it compares to the choices made by safetensors and dcp.

HF Safetensors (optimized for inference -- ~4GB file sizes, parallel lfs / xet io throughput)

Image

DCP (optimized for training -- rank specific files, low global metadata and resharding comms overhead)

Image

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions