o diy=@sdZddlZddlZddlmZddlmZmZmZm Z m Z ddl Z ddl m Z ddl mZddlmZddlmZdd lmZd d lmZd d lmZmZmZd d lmZddlmZmZddl m!Z!ddl"m#Z#ddl$m%Z%ddlm&Z&m'Z'erddl m(Z(e)Z*GdddZ+dS)z?Main class for performing graph optimization with ONNX Runtime.N)Path) TYPE_CHECKINGDictListOptionalUnion) load_model)GenerationConfig) AutoConfig) BertOnnxModel)optimize_model)check_model_uses_external_data) CONFIG_NAMENormalizedConfigManagerlogging)maybe_save_preprocessors)OptimizationConfig ORTConfig)ORTModelForCausalLM)ORTModel) ORTModelForConditionalGeneration)ONNX_WEIGHTS_NAMEORTConfigManager)PretrainedConfigc s,eZdZdZddeejdddeffdd Ze d d e e eje fd e ee d dfd dZ  d!dede e ejfde e de edef ddZede e ejfd ee effddZede e ejfde e ejfd efddZede e ejfde e ejfd ee effddZZS)" ORTOptimizerzc Handles the ONNX Runtime optimization process for models shared on huggingface.co/models. Fonnx_model_pathconfigr from_ortmodelcsbt||_||_|jj|_||_z t|j|j|_WdSt y0t d|jdw)a Args: onnx_model_path (`List[os.PathLike]`): The paths of the onnx models to optimize. config ([`~transformers.PretrainedConfig`]): An instance of the configuration associated to the model to optimize. from_ortmodel (`bool`, defaults to `False`): Whether the model being optimized is already loaded into an ORTModel, or if it was passed from disk. z-Tried to use ORTOptimizer for the model type zm, but it is not available yet. Please open an issue or submit a PR at https://github.com/huggingface/optimum.N) super__init__rr model_typerrget_normalized_config_classnormalized_configKeyErrorNotImplementedError)selfrrr __class__n/lsinfo/ai/hellotax_ai/llm_service/venv_embed/lib/python3.10/site-packages/optimum/onnxruntime/optimization.pyr!3s    zORTOptimizer.__init__N model_or_path file_namesreturncCsg}d}t|tr=d}t|tr'||jj|jjg7}|jdur&||jjnt|tr3|j r3t d||j|j }n?t j |rtd}|durLtgn|}t|}tt |vratdtdt|}|D] }|||qhntd|d||||dS) a Args: model_or_path (`Union[str, os.PathLike, ORTModel]`): The path to a local directory hosting the model to optimize or an instance of an `ORTModel` to quantize. Can be either: - A path to a local *directory* containing the model to optimize. - An instance of [`~optimum.onnxruntime.ORTModel`]. file_names(`Optional[List[str]]`, defaults to `None`): The list of file names of the models to optimize. NTaORTOptimizer does not support ORTModelForCausalLM models when without/with past models are merged. Please re-export your model. This can be done by using the optimum-cli ONNX export tool or `ORTModelForCausalLM.from_pretrained(..., export=True, use_merged=False)`.Fz= 2Gb. This argument is deprecated. one_external_file (`bool`, defaults to `True`): When `use_external_data_format=True`, whether to save all tensors to one external file. If False, save each tensor to a file named with the tensor name. Nz}The argument use_external_data_format in the ORTOptimizer.optimize() method is deprecated and will be removed in optimum 2.0.T)parentsexist_okzOptimizing model...F)load_external_data) optimizationrCrD_) opt_leveloptimization_optionsuse_gpuonly_onnxruntime)rCall_tensors_to_one_file)use_symbolic_shape_infer keep_io_typesz#Incomplete symbolic shape inferencezO. Try to set `disable_shape_inference=True` in your optimization configuration._datarzOptimized model saved at: z (external data format: z ; saved all tensor to one file: ))/loggerwarningrmkdirr"check_optimization_supported_modelr"get_model_ort_typercreate_fusion_optionsinforonnxloadstrrgccollectrr<stem with_suffixsuffixr as_posixr$num_attention_heads hidden_sizeoptimization_leveloptimize_for_gpu*enable_transformers_specific_optimizationsfp16save_model_to_filemodelconvert_float_to_float16disable_shape_inference Exception RuntimeErroris_filerr7removesave_pretrainedrparentr r;)r'r@rArBrCrDr"rLmodel_uses_external_data model_path onnx_model ort_configrb output_path optimizereerrgeneration_configr*r*r+optimizezs            zORTOptimizer.optimizecCsJtt|}|}tdddd|Ddd|DS)an Computes the dictionary mapping the name of the fused operators to their number of apparition in the model. Args: onnx_model_path (`Union[str, os.PathLike]`): Path of the ONNX model. Returns: The dictionary mapping the name of the fused operators to their number of apparition in the model. z%The following operators were fused : z, cSsg|] \}}|dkr|qSrr*.0kvr*r*r+ sz4ORTOptimizer.get_fused_operators..cSsi|] \}}|dkr||qSr~r*rr*r*r+ z4ORTOptimizer.get_fused_operators..)r rget_fused_operator_statisticsrTrZjoinitems)ronnx_optimized_modelfused_operatorr*r*r+get_fused_operatorss z ORTOptimizer.get_fused_operatorsonnx_optimized_model_pathcCsXtt|}tt|}t|}t|}||}td|d|d||S)a Compute the difference in the number of nodes between the original and the optimized model. Args: onnx_model_path (`Union[str, os.PathLike]`): Path of the ONNX model. onnx_optimized_model_path (`Union[str, os.PathLike]`): Path of the optimized ONNX model. Returns: The difference in the number of nodes between the original and the optimized model. z There are z nodes before optimization and z,nodes after. The number of nodes removed is )r rlennodesrTrZ)rrrvrnodes_number_onnx_model!nodes_number_onnx_optimized_modeldifference_nodes_numberr*r*r+get_nodes_number_differences    z(ORTOptimizer.get_nodes_number_differencecsxtt|tt|fddt}fD]}|D]}||jq qfdd|D}dd|DS)aN Compute the dictionary mapping the operators name to the difference in the number of corresponding nodes between the original and the optimized model. Args: onnx_model_path (`Union[str, os.PathLike]`): Path of the ONNX model. onnx_optimized_model_path (`Union[str, os.PathLike]`): Path of the optimized ONNX model. Returns: The dictionary mapping the operators name to the difference in the number of corresponding nodes between the original and the optimized model. cs$t|}t|}||SN)rget_nodes_by_op_type)op_typeonnx_model_nodes_with_op_type'onnx_optimized_model_nodes_with_op_type)rvrr*r+nodes_difference_given_typeAszJORTOptimizer.get_operators_difference..nodes_difference_given_typecsi|]}||qSr*r*)rr)rr*r+rLsz9ORTOptimizer.get_operators_difference..cSsi|] \}}|dkr||qSr~r*rr*r*r+rMr)r rsetraddrr)rrop_typesrknodeoperators_differencer*)rrvrr+get_operators_difference,s    z%ORTOptimizer.get_operators_difference)Fr)r?NT)__name__ __module__ __qualname____doc__rr7PathLikeboolr! classmethodrr]rrr;rr} staticmethodrintrrr __classcell__r*r*r(r+r.sZ"  2  &      r),rr^r7pathlibrtypingrrrrrr[r transformersr +transformers.models.auto.configuration_autor (onnxruntime.transformers.onnx_model_bertr "onnxruntime.transformers.optimizerr onnx.utilsrutilsrrrutils.save_utilsr configurationrrmodeling_decoderr modeling_ortrmodeling_seq2seqrrrr get_loggerrTrr*r*r*r+s.