o }oi@sdZddlZddlZddlZddlmZddlZddlm Z ddl m Z e Z Gddde Zejdd Zdd d Ze d dddZedge_dS)aHuggingFace model serialization support for NeMo's configuration system. This module provides integration between NeMo's configuration system and HuggingFace's pretrained models. It enables automatic serialization and deserialization of HuggingFace models within NeMo's configuration framework. The integration works by: 1. Detecting HuggingFace models through their characteristic methods (save_pretrained/from_pretrained) 2. Converting them to Fiddle configurations that preserve the model's class and path 3. Providing an artifact handler (HFAutoArtifact) that manages the actual model files Example: ```python from transformers import AutoModel # This model will be automatically handled by the HFAutoArtifact system model = AutoModel.from_pretrained("bert-base-uncased") # When serialized, the model files will be saved to the artifacts directory # When deserialized, the model will be loaded from the saved files ``` N)Path)Artifact) to_configc@s<eZdZdZdedededefddZdedefd d Zd S) HFAutoArtifacta4Artifact handler for HuggingFace pretrained model/processor/tokenizer/etc.. This handler manages the serialization and deserialization of HuggingFace models by utilizing their save_pretrained/from_pretrained methods. It saves models to an 'artifacts' subdirectory within the specified path. value absolute_dir relative_dirreturncCs&|t|ddtt|dS)aSave a HuggingFace model to disk. Args: instance: The HuggingFace model instance to save value: Original path value (unused) absolute_dir: Absolute path to the save directory relative_dir: Relative path from the config file to the save directory Returns: str: The relative path to the saved model artifacts artifactsz./)save_pretrainedrstr)selfinstancerrrrV/home/ubuntu/.local/lib/python3.10/site-packages/nemo/lightning/io/artifact/hf_auto.pydump<s zHFAutoArtifact.dumppathcCs|S)zReturn the path to load a HuggingFace model. Args: path: Path to the saved model artifacts Returns: Path: The same path, to be used with from_pretrained r)r rrrrloadKs zHFAutoArtifact.loadN)__name__ __module__ __qualname____doc__rrrrrrrr4srcksDttds it_tj}tj|z dVW|t_dS|t_w)a Context manager for passing additional kwargs to from_pretrained. Args: **kwargs: Keyword arguments to pass to from_pretrained Example: with from_pretrained_kwargs(trust_remote_code=True): io.load_context("path/to/checkpoint") kwargsN)hasattr_localrcopyupdate)rpreviousrrrfrom_pretrained_kwargsWs   rdummycCsttdi}|j|fi|S)aFactory function for loading HuggingFace pretrained models. This function is used as the serialization target for HuggingFace models. When deserialized, it will recreate the model using its from_pretrained method. Args: auto_cls: The HuggingFace model class (e.g., AutoModel, AutoTokenizer) pretrained_model_name_or_path: Path to the saved model or model identifier Returns: The loaded HuggingFace model r)getattrrfrom_pretrained)auto_clspretrained_model_name_or_pathrrrrr!ls r!cCs2t| ot|dddot|dot|dS)Nr transformersr r!)inspectisclassr startswithr)vrrr~s  r*cCstjt|jddS)aConvert a HuggingFace model instance to a Fiddle configuration. This handler detects HuggingFace model instances by checking for the presence of save_pretrained and from_pretrained methods. It converts them to a Fiddle configuration that will recreate the model using from_pretrained. Args: value: A HuggingFace model instance Returns: fdl.Config: A Fiddle configuration that will recreate the model r)r"r#)fdlConfigr! __class__)rrrrhandle_hf_pretrained}s r.r#)r)r contextlibr& threadingpathlibrfiddler+nemo.lightning.io.artifactrnemo.lightning.io.to_configrlocalrrcontextmanagerrr!registerr.__io_artifacts__rrrrs$   #