[Feature] Distributed model decompression

[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.

[Vishnu-sai-teja]

@kylesayrs I can take this up !

[kylesayrs]

@Vishnu-sai-teja Feel free! The relevant logic is here

[Vishnu-sai-teja]

Thank you !