o TÃibã@sŒddlmZddlmZmZmZmZddlZddlm Z ddl m Z ddl m Z ddlmZd d lmZGd d „d e ƒZGd d„de ƒZdS)é)Úabstractmethod)ÚAnyÚDictÚOptionalÚTypeN)ÚDeepSpeedConfigModelé)Ú DSModuleBase)ÚDSModuleRegistryBase©ÚDSLinearConfigé)ÚInferenceParameterc s¤eZdZdZedeefdd„ƒZdede e e fddf‡fdd „ Z e d ejdefd d „ƒZdd ejdejdeejdejfdd„Zee dejfdd„ƒƒZ‡ZS)Ú DSLinearBasea& Base mixin for all Linear modules. The interface represented by this module is: hidden_out = activation(hidden_in * weight + bias) The format and dtype of the weight and bias tensors are not defined and implementations may compress as necessary. Must support a bias. ÚreturncCótS©Nr ©rrúi/home/ubuntu/.local/lib/python3.10/site-packages/deepspeed/inference/v2/modules/interfaces/linear_base.pyÚ config_classózDSLinearBase.config_classÚconfigÚimplementation_configNcstƒ ||¡dSr)ÚsuperÚ__init__)Úselfrr©Ú __class__rrr!szDSLinearBase.__init__ÚparamcCódS)z§ Perform any necessary transformations of the parameters of this module. Parameters: param (torch.Tensor): Weight or bias tensor. Nr)rrrrrÚtransform_param$szDSLinearBase.transform_paramÚ hidden_statesÚwÚbcCstƒ‚)aF Parameters: hidden_states (torch.Tensor): Hidden states tensor. Expected shape is either [batch, seq_len, in_channels] or [batch, in_channels]. Returns: torch.Tensor: Output tensor. Tensor should have same number of dimensions as input tensor. )ÚNotImplementedError)rr!r"r#rrrÚforward.s zDSLinearBase.forwardcCr)zA Return the padded, pre-allocated output Tensor. Nr)rrrrÚoutput:szDSLinearBase.outputr)Ú__name__Ú __module__Ú __qualname__Ú__doc__Ú staticmethodrrrr rÚstrrrrÚtorchÚTensorrr rr%Úpropertyr&Ú __classcell__rrrrrs "(  rc@s0eZdZUiZeed<edeefdd„ƒZ dS)ÚDSLinearRegistryÚregistryrcCrr)rrrrrÚassociated_classFrz!DSLinearRegistry.associated_classN) r'r(r)r2rÚ__annotations__r+rr r3rrrrr1Cs r1)ÚabcrÚtypingrrrrr-Údeepspeed.runtime.config_utilsrÚ ds_moduler Úmodule_registryr Úconfigsr Úinference_parameterrrr1rrrrÚs      1