pymllm.layers.base

Classes

MllmBaseLayer

Module Contents

class pymllm.layers.base.MllmBaseLayer

Bases: torch.nn.Module

quant_method: pymllm.layers.quantize_base.QuantizeMethodBase | None = None
weight_loader(param, loaded_weight)

Load weights into a parameter.

This is the default implementation that directly copies the loaded weight into the parameter. Subclasses should override this method to implement custom loading logic (e.g., tensor parallelism sharding).

Parameters:

  • param (torch.nn.Parameter) – The parameter to load weights into.

  • loaded_weight (torch.Tensor) – The weight tensor loaded from checkpoint.

abstractmethod forward(*args, **kwargs)