Purpose

• Refactor module targeting to be cleaner and easier to maintain
• Support skipping TransformBase modules

Prerequisites

• Deprecate iter_named_leaf_modules and iter_named_quantizable_modules vllm-project/llm-compressor#1628

Changes

• Deprecated iter_named_leaf_modules and iter_named_quantizable_modules
  • These functions are near copies of each other
  • These functions don't actually have any runtime savings over simple model.modules(). Infact, they're likely significantly slower
  • These functions hard code against "observer" module names and are difficult to extend to other modules like Transforms
• Add UntargetableModule class

class UntargetableModule(torch.nn.Module):
    """
    Abstract base class for modules which are not a part of the the model definition.
    `torch.nn.Module`s which inherit from this class will not be targeted by configs
    This is typically used to skip apply configs to `Observers` and `Transforms`
    """

    pass

• Observer and TransformBase inherit from this class

TODO

• Add skip in transforms apply

Testing

• https://github.com/neuralmagic/llm-compressor-testing/actions/runs/16123598340 ✅