Skip to content

[Feature] Distributed model decompression #671

Description

@kylesayrs

Background

#624 Added support for distributed weight compression which parallelizes weight compression for distributed workflows. However, weight decompression has not been parallelized, meaning that it can take a long time to decompress a model during transformers inference or use cases where a user wants to decompress a model.

Requested Changes

Implement distributed decompression (ModelCompressor.decompress_model). This involves supporting BaseCompressor.decompress_module(module) on modules whose parameters are on the meta device for all subclasses of BaseCompressor.

Please add tests to verify that BaseCompressor.decompress_module and BaseCompressor.compress_module work for meta modules for all subclasses.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

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