[Hotfix] Implement quantization compressor methods on dense compressor

[kylesayrs]

Background

• The typing of ModelCompressor.quantization_compressor is less narrow than I expected when implementing [Performance] Add memory compression and decompression pathways #301

Purpose

• Fix model decompression in cases where the quantization compressor is a dense compressor
  • [Examples] [Bugfix] skip sparsity stats when saving checkpoints vllm-project/llm-compressor#1528

Changes

• Implement decompress_module_from_state_dict on DenseCompressor

[rahul-tuli]

Should we just delete the dense compressor?

[kylesayrs]

@rahul-tuli We can talk more about this somewhere else. I'd rather keep the/a dense compressor and get rid of the nullable property of compressors. This would make a lot of code much cleaner.

if compressor is not None:
  compressor.compress(model)

if compressor is not None:
  compressor.decompress(model)

is a lot more annoying to check every time you want to use a compressor programmatically, rather than a simply narrow type

compressor.compress(model)
compressor.decompress(model)

[rahul-tuli]

@rahul-tuli We can talk more about this somewhere else. I'd rather keep the/a dense compressor and get rid of the nullable property of compressors. This would make a lot of code much cleaner.

if compressor is not None:

  compressor.compress(model)



if compressor is not None:

  compressor.decompress(model)

is a lot more annoying to check every time you want to use a compressor programmatically, rather than a simply narrow type

compressor.compress(model)

compressor.decompress(model)

Sounds good @kylesayrs, could you create a ticket/issue so we remember to revisit this?

Approving the PR!

[shanjiaz]

LGTM