GPTQ Algorithm Cleanup

[kylesayrs]

Purpose

1. Clean up implementation for easier reading (comments, better structure)
2. Allow the algorithm to be skipped if the layer is not being targeted
3. Fix bug where layer is not frozen after QuantizationModifier
4. Prevent weight observer misuse
5. Depreciate weight_fake_quant use case

Changes

• ensure that freeze_quantization is True (default), even if QuantizationModifier is wrapped by GPTQModifier
• implement get_attr_chain helper function to be used for getting chained attributes
• use get_attr_chain to get weight quantization arguments and skip computation if weight does not have valid args
• directly use memoryless observer to avoid misuse with unsupported observers
• perform transpose and float conversion in place to reduce memory use
• break out logging operations to separate function
• remove weight_fake_quant cases

Testing

Regression tested saving, loading, and vllm inferencing with group quantized model

[kylesayrs]

@Satrat Can you specify what you're looking for in a skip test?

[Satrat]

@Satrat Can you specify what you're looking for in a skip test?

You could just initialize a module with some modules skipped (more than the lm_head) and others quantized, then search the logs for the debug string, or just testing your getattr_chain helper function directly on the model would be fine too

[kylesayrs]

Yeah the failing base test is because of a bug from the previous release which I fixed in the main branch See: https://github.com/neuralmagic/compressed-tensors/blame/4b214e582c8434733efea79239cfadec9358b7fb/src/compressed_tensors/quantization/observers/base.py#L165-L167

[kylesayrs]

Using my local machine and the main branch of compressed_tensors, I confirmed that the tests/llmcompressor/modifiers/ and tests/llmcompressor/transformers/compression/ are passing