o ߥi/@sddlZddlmZddlmZddlmZmZmZm Z m Z m Z ddl Z ddl mZddl mZddlmZmZddlmZmZmZdd lmZdd lmZd d lmZeZGd ddee jjZdS)N)deepcopy)partial)AnyCallableDictListOptionalUnion)version)nn) DataParallelDistributedDataParallel)save_checkpointsave_configurationsave_pretrained)func_receive_dict_inputs) get_logger)Modelc seZdZdZdfdd ZdeeeffddZ  d d d Z deeeffd d Z ddZ ddZ ddZ deedddefdeeejfdeeeefdedeedef ddZddZZS)! TorchModelz' Base model interface for pytorch Ncs,tj|g|Ri|tjj|dSN)super__init__torchr Module)self model_dirargskwargs __class__[/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/base/base_torch_model.pyrszTorchModel.__init__returncOs<t|jr||j|dfi|S||j|i|S)Nr)rforward postprocessrrrr!r!r"__call__s zTorchModel.__call__Tparamsc Cst|ttfr |j}tj|ddd}|dur2||vr*d|vr*d}td|d||vr2||}td|jj d|d |d t | D]\}}| d r_|||d d<| |qI|j||d td|S)NcSs|Srr!)storagelocr!r!r".sz-TorchModel._load_pretrained..) map_locationr(z Loading: z does not exist, use params.zLoading z model from z, with param key: [z].zmodule.)strictzload model done.) isinstancer r modulerloadloggerinfor __name__ritems startswithpopload_state_dict)rnet load_pathr. param_keyload_netkvr!r!r"_load_pretrained&s0    zTorchModel._load_pretrainedcOstr)NotImplementedErrorr&r!r!r"r$BszTorchModel.forwardcCs |dS)z A method executed at the end of each model initialization, to execute code that needs the model's modules properly initialized (such as weight initialization). N) init_weightsrr!r!r" post_initEs zTorchModel.post_initcCs||jdSr)apply _init_weightsrBr!r!r"rALszTorchModel.init_weightscCst|tjr|jjjddd|jdur|jjdSdSt|tjr?|jjjddd|j dur=|jj|j dSdSt|tj rT|jj|jj ddSdS)zInitialize the weightsgg{Gz?)meanstdNg?) r/r Linearweightdatanormal_biaszero_ Embedding padding_idx LayerNormfill_)rr0r!r!r"rEPs      zTorchModel._init_weightsF) with_meta target_foldersave_checkpoint_names save_functionconfigsave_config_functioncKsH|dur t|dr |j}t||||fi||dur"|||dSdS)asave the pretrained model, its configuration and other related files to a directory, so that it can be re-loaded Args: target_folder (Union[str, os.PathLike]): Directory to which to save. Will be created if it doesn't exist. save_checkpoint_names (Union[str, List[str]]): The checkpoint names to be saved in the target_folder save_function (Callable, optional): The function to use to save the state dictionary. config (Optional[dict], optional): The config for the configuration.json, might not be identical with model.config save_config_function (Callble, optional): The function to use to save the configuration. Ncfg)hasattrrXr)rrSrTrUrVrWrr!r!r"r`szTorchModel.save_pretrainedcKs@ttjtdkrtj|fi|Stdtjd|S)zCompile torch model with torch>=2.0 Args: kwargs: backend: The backend param of torch.compile mode: The mode param of torch.compile z 2.0.0.devzFTorch compiling needs torch version >= 2.0.0, your torch version is : z, returns original model)r parser __version__compiler2warning)rrr!r!r"r\s  zTorchModel.compiler)Tr()r4 __module__ __qualname____doc__rrstrrr'r?r$rCrArErrrr osPathLikerrrdictrr\ __classcell__r!r!rr"rs:    %r) rbcopyr functoolsrtypingrrrrrr r packagingr r torch.nn.parallelr r modelscope.utils.checkpointrrrmodelscope.utils.file_utilsrmodelscope.utils.loggerr base_modelrr2rrr!r!r!r"s