[Docs] GPTQ Docstring, better argument grouping

[kylesayrs]

Background

#36 Updated the GPTQModifier argument name, but did not change the docstring

Changes

• Fix targets docstring on GPTQModifier
• Rearrange modifier arguments for clarity

Testing

• N/A

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

[dsikka]

Is the targets param not used anywhere? seems like the last PR kept it, on line 101

[kylesayrs]

@dsikka targets is used, but in the sense that it is an argument for QuantizationModifier which piggybacks on GPTQModifier. The docstring only includes GPTQModifier-specific arguments. This "piggybacking" design will be removed in the lifecycle refactor.

[markurtz]

Why are renaming / standardizing on sequential targets? The original goal, across all modifiers, was to standardize on targets as the base name to represent the scope of which the modifier would target -- I'd prefer not to deviate from that whenever we can.

[kylesayrs]

@markurtz The GPTQModifier has two targets arguments, targets which selects which modules to compress, and sequential_targets, which selects which modules are considered layers (transformer blocks) for the purposes of true_sequential=False. Both are required, since the inferred defaults are not always correct and should be capable of being overridden.

[kylesayrs]

Closing for #914