S^(hQdZddlZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddl Z ddlZddlZddlZddlZddlmZddlmZddlmZmZmZmZmZddlmZddlmZ ddl!Z"ddl#Z#ddl$m%Z&ddl'm(Z(m)Z)m*Z*dd l+m,Z,dd l#m-Z-dd l.m/Z/m0Z0m1Z1m2Z2m3Z3dd l4m5Z5dd l6m7Z7ddl8m9Z9m:Z:m;Z;ddlZ>ddl?m@Z@ddlAmBZBddlCmDZDmEZEddlFmGZGddlHmIZImJZJmKZKddlLmMZMddlNmOZOddlPmQZQmRZRmSZSddlTmUZUmVZVddlWmXZXmYZYddlZm[Z[ddl\m]Z]m^Z^ddl_m`Z`ddlambZbmcZcmdZdmeZemfZfmgZgmhZhmiZiddljmkZkmlZlmmZmmnZnmoZompZpmqZqmrZrmsZsmtZtmuZumvZvmwZwmxZxmyZymzZzm{Z{m|Z|m}Z}m~Z~ddlmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZdd lmZmZmZdd!lmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZdd"lmZdd#lmZecgZefZerdd$lmZeZerdd%lmZerddlZerQddlmcmZddlmcmZdd lm5Ze,jeثe,jek\Zerddlm%cmZddlmZnd&ZerCddlmcm#Zdd lm5Ze,jee,jd'k\Zdd(ljmZmZmZmZnd&ZerddlZerdd)lmZerdd*lmZmZdd lm5Zdd+lmZdd,lmZmZmZmZmZmZmZe2gZe,jee,jd-kDrdd.lmZe,jee,jd/kDrdd0lmZeegz ZeKrdd1lmZed2rdd3lmZd4Zd5Zd6ZerddlZerddlZejeZ d7Z d8Z d9Z d:Z d;Zd<Zd=ZGd>d?Zy)@uc The Trainer class, to easily train a 🤗 Transformers from scratch or finetune it on a new task. N)Mapping)Path) TYPE_CHECKINGAnyCallableOptionalUnion)#get_reporting_integration_callbacks) ModelCard create_repo upload_folder)version)nn) DataLoaderDatasetIterableDataset RandomSamplerSequentialSampler) __version__)PretrainedConfig) DataCollatorDataCollatorWithPaddingdefault_data_collator) DebugOptionDebugUnderflowOverflow)SequenceFeatureExtractor)FeatureExtractionMixin)"ALL_HYPERPARAMETER_SEARCH_BACKENDSdefault_hp_search_backend)BaseImageProcessor)deepspeed_initdeepspeed_load_checkpointis_deepspeed_available)tpu_spmd_dataloader)TrainingSummary)PreTrainedModelload_sharded_checkpoint unwrap_model)!MODEL_FOR_CAUSAL_LM_MAPPING_NAMESMODEL_MAPPING_NAMES) Adafactor get_scheduler)ProcessorMixin)ALL_LAYERNORM_LAYERS"is_torch_greater_or_equal_than_2_3)PreTrainedTokenizerBase)CallbackHandlerDefaultFlowCallbackExportableStatePrinterCallbackProgressCallbackTrainerCallbackTrainerControl TrainerState)DistributedTensorGathererEvalLoopContainerIterableDatasetShard LabelSmootherLayerWiseDummyOptimizerLengthGroupedSamplerSequentialDistributedSamplerdistributed_broadcast_scalarsdistributed_concatfind_batch_sizeget_model_param_countget_module_class_from_nameget_parameter_names nested_concat nested_detachnested_numpifynested_xla_mesh_reducereissue_pt_warningsremove_dummy_checkpointset_rng_state_for_device)PREFIX_CHECKPOINT_DIRBestRunEvalLoopOutputEvalPredictionHPSearchBackend HubStrategyPredictionOutputRemoveColumnsCollator SaveStrategyTrainerMemoryTracker TrainOutputcheck_target_module_existsdefault_compute_objectivedenumpify_detensorizeenable_full_determinismfind_executable_batch_sizeget_last_checkpoint has_lengthneftune_post_forward_hooknumber_of_arguments seed_workerset_seed speed_metrics)OptimizerNames ParallelModeTrainingArguments))ADAPTER_CONFIG_NAMEADAPTER_SAFE_WEIGHTS_NAMEADAPTER_WEIGHTS_NAME CONFIG_NAMESAFE_WEIGHTS_INDEX_NAMESAFE_WEIGHTS_NAMEWEIGHTS_INDEX_NAME WEIGHTS_NAMEXLA_FSDPV2_MIN_VERSIONPushInProgressPushToHubMixincan_return_loss find_labelsis_accelerate_availableis_apex_availableis_apollo_torch_availableis_bitsandbytes_availableis_datasets_availableis_galore_torch_availableis_grokadamw_availableis_in_notebookis_ipex_availableis_liger_kernel_availableis_lomo_availableis_peft_availableis_safetensors_availableis_sagemaker_dp_enabledis_sagemaker_mp_enabledis_schedulefree_availableis_torch_compile_availableis_torch_hpu_availableis_torch_mlu_availableis_torch_mps_availableis_torch_musa_availableis_torch_neuroncore_availableis_torch_npu_availableis_torch_xla_availableis_torch_xpu_availableis_torchao_availablelogging strtobool)deprecate_kwarg)QuantizationMethod)NotebookProgressCallback)ampFz1.10)smp_forward_backwardsmp_forward_only smp_gathersmp_nested_concat) PeftModel) Acceleratorskip_first_batches)AcceleratorState)AutocastKwargsDistributedDataParallelKwargsDistributedTypeload_fsdp_modelload_fsdp_optimizersave_fsdp_modelsave_fsdp_optimizer1.3.0)TorchTensorParallelPluginz0.23.0)SeedableRandomSampler)DeepSpeedSchedulerWrapper0.28.0)DataLoaderConfigurationctrttrtfnd}tjtj jdtjdk\r ddlm}g||}t||Sy)Npeftz0.7.0r)PeftMixedModelF) rrrparse importlibmetadatarr isinstance)modelclasses_to_checkrs R/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/transformers/trainer.py_is_peft_modelrsg+<+>IB %!122 c~tr2dttjtj vrddiSiS)N adapter_onlyT)rulistinspect signaturer parametersrrr_get_fsdp_ckpt_kwargsrs5 ^tGdeefd/Z?d0Z@d1ZAd2ZBdd3ee eejjjffd4ZEeF dd ed ee dee%e%ffd5ZGdd-e=d6ej@jBfd7ZHd8e8de=fd9ZIeFdd:e8d;ee=de=fd<ZJd=e d>eee%fffd?ZKd=e d>eee%ffd@e=dAeeeLffdBZMdCefdDZNddEZOddGZPdFejfdHZRdIZSddJZT ddKee eeUfd=e d>eee%fd fdLeeefdMZV ddNZWdOZXddPZYdQZZdRZ[ddSZ\ ddTZ]dUZ^dVZ_dWZ`dXZadYZbdZZcd[Zdd\Zed]Zf dd^eed>geeeLffd_eeeeeLfgeLfd`e=dae eeefdbee dcegfddeed>gefde eheehffdeZiddfeeeLfdgeeLdd fdhZjdie ejLe%fde ejLe%ffdjZkdkeee ejLe%ffdeee ejLe%fffdlZldmZmddneeUfdoZn dd e jdkeee ejLe%ffdejLfdpZoddqZpdeUfdrZqdeUfdsZrddteedueUfdvZsddteefdwZtd dteefdxZudyZvd ewdFfdeefdzZxddd{Zy ddee eeeeffd|eeed}edeeeLffd~Zz dd+ed|eeed}ede{fdZ| dd8e8d$edeeUd|eeed}ede}f dZ~ddZ dd e jdkeee ejLe%ffdeUd|eeedeeejLeejLeejLff dZdkeee ejLe%fffdZddeefdZ ddeedeede eeed fdeedeede eeed fde eeed fd#e eeed fde eeed ffdZdZdZ ddeedeUdeedeedef dZ dd8e8d$edeeUd|eeed}ede}f dZdZddZdZddZdZdZd ed8e8de=fdZy )TraineruA Trainer is a simple but feature-complete training and eval loop for PyTorch, optimized for 🤗 Transformers. Args: model ([`PreTrainedModel`] or `torch.nn.Module`, *optional*): The model to train, evaluate or use for predictions. If not provided, a `model_init` must be passed. [`Trainer`] is optimized to work with the [`PreTrainedModel`] provided by the library. You can still use your own models defined as `torch.nn.Module` as long as they work the same way as the 🤗 Transformers models. args ([`TrainingArguments`], *optional*): The arguments to tweak for training. Will default to a basic instance of [`TrainingArguments`] with the `output_dir` set to a directory named *tmp_trainer* in the current directory if not provided. data_collator (`DataCollator`, *optional*): The function to use to form a batch from a list of elements of `train_dataset` or `eval_dataset`. Will default to [`default_data_collator`] if no `processing_class` is provided, an instance of [`DataCollatorWithPadding`] otherwise if the processing_class is a feature extractor or tokenizer. train_dataset (Union[`torch.utils.data.Dataset`, `torch.utils.data.IterableDataset`, `datasets.Dataset`], *optional*): The dataset to use for training. If it is a [`~datasets.Dataset`], columns not accepted by the `model.forward()` method are automatically removed. Note that if it's a `torch.utils.data.IterableDataset` with some randomization and you are training in a distributed fashion, your iterable dataset should either use a internal attribute `generator` that is a `torch.Generator` for the randomization that must be identical on all processes (and the Trainer will manually set the seed of this `generator` at each epoch) or have a `set_epoch()` method that internally sets the seed of the RNGs used. eval_dataset (Union[`torch.utils.data.Dataset`, Dict[str, `torch.utils.data.Dataset`, `datasets.Dataset`]), *optional*): The dataset to use for evaluation. If it is a [`~datasets.Dataset`], columns not accepted by the `model.forward()` method are automatically removed. If it is a dictionary, it will evaluate on each dataset prepending the dictionary key to the metric name. processing_class (`PreTrainedTokenizerBase` or `BaseImageProcessor` or `FeatureExtractionMixin` or `ProcessorMixin`, *optional*): Processing class used to process the data. If provided, will be used to automatically process the inputs for the model, and it will be saved along the model to make it easier to rerun an interrupted training or reuse the fine-tuned model. This supercedes the `tokenizer` argument, which is now deprecated. model_init (`Callable[[], PreTrainedModel]`, *optional*): A function that instantiates the model to be used. If provided, each call to [`~Trainer.train`] will start from a new instance of the model as given by this function. The function may have zero argument, or a single one containing the optuna/Ray Tune/SigOpt trial object, to be able to choose different architectures according to hyper parameters (such as layer count, sizes of inner layers, dropout probabilities etc). compute_loss_func (`Callable`, *optional*): A function that accepts the raw model outputs, labels, and the number of items in the entire accumulated batch (batch_size * gradient_accumulation_steps) and returns the loss. For example, see the default [loss function](https://github.com/huggingface/transformers/blob/052e652d6d53c2b26ffde87e039b723949a53493/src/transformers/trainer.py#L3618) used by [`Trainer`]. compute_metrics (`Callable[[EvalPrediction], Dict]`, *optional*): The function that will be used to compute metrics at evaluation. Must take a [`EvalPrediction`] and return a dictionary string to metric values. *Note* When passing TrainingArgs with `batch_eval_metrics` set to `True`, your compute_metrics function must take a boolean `compute_result` argument. This will be triggered after the last eval batch to signal that the function needs to calculate and return the global summary statistics rather than accumulating the batch-level statistics callbacks (List of [`TrainerCallback`], *optional*): A list of callbacks to customize the training loop. Will add those to the list of default callbacks detailed in [here](callback). If you want to remove one of the default callbacks used, use the [`Trainer.remove_callback`] method. optimizers (`Tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`, *optional*, defaults to `(None, None)`): A tuple containing the optimizer and the scheduler to use. Will default to an instance of [`AdamW`] on your model and a scheduler given by [`get_linear_schedule_with_warmup`] controlled by `args`. optimizer_cls_and_kwargs (`Tuple[Type[torch.optim.Optimizer], Dict[str, Any]]`, *optional*): A tuple containing the optimizer class and keyword arguments to use. Overrides `optim` and `optim_args` in `args`. Incompatible with the `optimizers` argument. Unlike `optimizers`, this argument avoids the need to place model parameters on the correct devices before initializing the Trainer. preprocess_logits_for_metrics (`Callable[[torch.Tensor, torch.Tensor], torch.Tensor]`, *optional*): A function that preprocess the logits right before caching them at each evaluation step. Must take two tensors, the logits and the labels, and return the logits once processed as desired. The modifications made by this function will be reflected in the predictions received by `compute_metrics`. Note that the labels (second parameter) will be `None` if the dataset does not have them. Important attributes: - **model** -- Always points to the core model. If using a transformers model, it will be a [`PreTrainedModel`] subclass. - **model_wrapped** -- Always points to the most external model in case one or more other modules wrap the original model. This is the model that should be used for the forward pass. For example, under `DeepSpeed`, the inner model is wrapped in `DeepSpeed` and then again in `torch.nn.DistributedDataParallel`. If the inner model hasn't been wrapped, then `self.model_wrapped` is the same as `self.model`. - **is_model_parallel** -- Whether or not a model has been switched to a model parallel mode (different from data parallelism, this means some of the model layers are split on different GPUs). - **place_model_on_device** -- Whether or not to automatically place the model on the device - it will be set to `False` if model parallel or deepspeed is used, or if the default `TrainingArguments.place_model_on_device` is overridden to return `False` . - **is_in_train** -- Whether or not a model is currently running `train` (e.g. when `evaluate` is called while in `train`) r )_get_learning_rate log_metricsmetrics_format save_metrics save_state tokenizerprocessing_classz5.0.0T)new_namerraise_if_both_namesNNNrargs data_collator train_datasetzdatasets.Dataset eval_dataset model_initcompute_loss_funccompute_metrics callbacks optimizersoptimizer_cls_and_kwargspreprocess_logits_for_metricsc v |'d}tjd|dt|}|jr<| :dt j | j jvr td|j*|jdk7r|td|jd |jtjk(s |jr|j td ||_||_|jj"rt%|jj&nt)|jj&d|_d|_d |_|j1t3|jj4|_|j6j9|j;}t=j>||j@|%|||_!|jE}n.tGd |tIjJd tL||_!|jNjPtRvr#td|jNjPdtU|dd rtU|dd rd|_+nd |_+tU|ddtY|jZj]Dcgc] }|dvs| }}t_|dkDrd|_+nJt_|dk(r5|jj`tcj`|dk7|_+nd |_+|jVrtjd|jjdrtgrxddl4m5}tm|tnr ||nctq|dr6tm|jstnr||jsn!tjudn twdtU|dd xrtU|dd  }tU|ddduxr|jxjz}tU|ddduxrtU|jxd d }|rtq|d!r td"|rt}|s |s td#|rO|sMtd$|jxj~jd%|jxj~j|jd&|_Bt_|jdkDrN|jr td'|jd&s(|jtjk7r td(|j|_H|jVsH|js<|js |jr |jr|js |jrd |_H|!tm|ttfr t|nt}||n||_Q||_R||_S||_T|jr:tU|d)dtjk(s|j||j`|jVrd|j_X||_Y||_Z|jj|}t}|s |jn|jsj}t j |j }tq|d*r|j|__n%td+|j]D|__|j|_a| |_b| |_c| \|_d|_e| |_f|j|j tGd,|#|j |j tGd-tr|j||jj D]}|j`}n|jjD]'}t_|d.dkDs|d.dj`}nk7r td/|js |jr#|j |j tGd0tt|jjz}| |n|| z} t| |j|j|j|j|_m|j|jjrtntd |_rd|_s|jjr|j|jjr+tj|jjd1t|js+ttU|jd2dr td3|jdkDr$|jdkDrtjd4|%t|s|jdkr td5|Gtm|tbjjjr|jr td6d|_d |_d |_t r_|j r td7tr|jtjjjk7r tjud8tjjjd9|jd:tjjjtjjj|_nftqtjjd;r?tjud8tjjjd<|js |j r||jd=k(rl|j`tcj`d>k(r&|jrtstd?d@|_tjdA|jdB|js |j rr|jsft s[|jd@k(rd|_tbj|_n-|jdCk(rt!s twdDd|_|jj"dk7r(t%|jj"E|_nd|_t)|_t-|j/|j1|jj2|j*gzDcgc]}tm|t4s|c}F|_d|_d|_t}|jrN|jj:7tjudG|jjNjPdHt=|jjN} |jj:| n|jj:|_t?|jjN|_|jjA|j|j|j*|_|jB|_d |_|j6jI|jJrtMs tGdI|jjOdJd |_|jPrxtRs tdKtUjV}!tYjZtYj\t_j`tc|!|!dfdLM|jxr|jP |_ycc}wcc}w)NN tmp_trainerz1No `TrainingArguments` passed, using `output_dir=z`.) output_dircompute_resultzWhen using `batch_eval_metrics`, your `compute_metrics` function must take a `compute_result` boolean argument which will be triggered after the last batch of the eval set to signal that the summary statistics should be returned by the function.noz%You have set `args.eval_strategy` to zx but you didn't pass an `eval_dataset` to `Trainer`. Either set `args.eval_strategy` to `no` or pass an `eval_dataset`. z`args.metric_for_best_model` must be provided when using 'best' save_strategy or if `args.load_best_model_at_end` is set to `True`.Fz<`Trainer` requires either a `model` or `model_init` argumentz`Trainer` requires either a `model` or `model_init` argument, but not both. `model_init` will overwrite your model when calling the `train` method. This will become a fatal error in the next release.zThe model you have picked (a) cannot be used as is for training: it only computes hidden states and does not accept any labels. You should choose a model with a head suitable for your task like any of the `AutoModelForXxx` listed at https://huggingface.co/docs/transformers/model_doc/autois_parallelizablemodel_parallelT hf_device_map)cpudiskr rzYou have loaded a model on multiple GPUs. `is_model_parallel` attribute will be force-set to `True` to avoid any unexpected behavior such as device placement mismatching.)_apply_liger_kernel_to_instancerget_base_modelzRThe model is not an instance of PreTrainedModel. No liger kernels will be applied.zYou have set `use_liger_kernel` to `True` but liger-kernel >= 0.3.0 is not available. Please install it with `pip install liger-kernel` is_quantized_hf_peft_config_loaded hf_quantizeris_qat_trainable _orig_modzYou cannot fine-tune quantized model with `torch.compile()` make sure to pass a non-compiled model when fine-tuning a quantized model with PEFTzYou cannot perform fine-tuning on purely quantized models. Please attach trainable adapters on top of the quantized model to correctly perform fine-tuning. Please see: https://huggingface.co/docs/transformers/peft for more detailsz8The model you are trying to fine-tune is quantized with z but that quantization method do not support training. Please open an issue on GitHub: https://github.com/huggingface/transformers to request the support for training support for xlazZUsing --fsdp xxx together with --deepspeed is not possible, deactivate one of those flags.z.Using fsdp only works in distributed training.quantization_methodaccepts_loss_kwargsc3jK|]+}|jtjjk(-ywN)kindr Parameter VAR_KEYWORD).0ks r z#Trainer.__init__..~s*1<='++7771s13zSPassing both `optimizers` and `optimizer_cls_and_kwargs` arguments is incompatible.zPassing a `model_init` is incompatible with providing the `optimizers` argument. You should subclass `Trainer` and override the `create_optimizer_and_scheduler` method.paramsa[The model and the optimizer parameters are not on the same device, which probably means you created an optimizer around your model **before** putting on the device and passing it to the `Trainer`. Make sure the lines `import torch_xla.core.xla_model as xm` and `model.to(xm.xla_device())` is performed before the optimizer creation in your script.zPassing `optimizers` is not allowed if PyTorch FSDP is enabled. You should subclass `Trainer` and override the `create_optimizer_and_scheduler` method.exist_ok collate_batchzRThe `data_collator` should be a simple callable (function, class with `__call__`).zHmax_steps is given, it will override any value given in num_train_epochszThe train_dataset does not implement __len__, max_steps has to be specified. The number of steps needs to be known in advance for the learning rate scheduler.zTthe `--group_by_length` option is only available for `Dataset`, not `IterableDatasetzOSageMaker Model Parallelism does not support BF16 yet. Please use FP16 instead z(FP16 provided in SM_HP_MP_PARAMETERS is z+, but FP16 provided in trainer argument is z , setting to fp16zJ, but SageMaker Model Parallelism < 1.10 does not support FP16 in trainer.autorz2Tried to use `fp16` but it is not supported on cpucpu_ampzUsing z half precision backendapexzcUsing FP16 with APEX but APEX is not installed, please refer to https://www.github.com/nvidia/apex.)epsilon)is_local_process_zerois_world_process_zerostateful_callbacksz)No label_names provided for model class `z`. Since `PeftModel` hides base models input arguments, if label_names is not given, label_names can't be set automatically within `Trainer`. Note that empty label_names list will be used instead.z3Using torch.compile requires PyTorch 2.0 or higher. xla_fsdp_v2z*FSDPv2 requires `torch_xla` 2.2 or higher.)fsdptensor) axis_names)loggerinforgbatch_eval_metricsrrrkeys ValueError eval_strategy save_strategyrVBESTload_best_model_at_endmetric_for_best_modelrrfull_determinismr\seedrchp_name deepspeed is_in_train"create_accelerator_and_postprocessrWskip_memory_metrics_memory_trackerstartget_process_log_levelr set_verbosity_setup_devicesrcall_model_init RuntimeErrorwarningswarn FutureWarning __class____name__r+getattris_model_parallelsetrvalueslendeviceruse_liger_kernelr~liger_kernel.transformersrrr'hasattrrwarning ImportErrorr is_trainablerquantization_config quant_method fsdp_configis_fsdp_xla_enabledris_deepspeed_enabled parallel_moderf DISTRIBUTEDplace_model_on_devicefp16_full_evalbf16_full_evaldo_trainis_fsdp_enabledr1rrrrrrrrBITS_AND_BYTES_move_model_to_device_n_gpu model_wrappedr acceleratorr)forwardrmodel_accepts_loss_kwargsanyneftune_noise_alpharr optimizer lr_schedulerrr param_groupsDEFAULT_CALLBACKSr report_tor2callback_handler add_callback disable_tqdmr5DEFAULT_PROGRESS_CALLBACK_loggers_initialized hub_model_id push_to_hub init_hf_repo should_saveosmakedirsrcallable max_stepsnum_train_epochsr_utilsdatargroup_by_length_signature_columnsuse_apex use_cpu_amprbf16IS_SAGEMAKER_MP_POST_1_10r smpstatecfghalf_precision_backendr0bfloat16 amp_dtypervlabel_smoothing_factorr=label_smootherr8controlr9rrrr4 current_floshp_search_backend label_namesrtrs on_init_endtrain_batch_size_train_batch_size_created_lr_schedulerstop_and_update_metrics torch_compilergetis_fsdp_xla_v2_enabledIS_XLA_FSDPV2_POST_2_2xrglobal_runtime_device_countxsset_global_meshMeshrarrayrangeis_fsdp_xla_v1_enabled)"selfrrrrrrrrrrrrrr log_levelr8devicesr_is_quantized_and_base_model&_quantization_method_supports_training%_is_model_quantized_and_qat_trainabledefault_collatorunwrapped_model model_forwardforward_paramsparam model_device param_groupoptimizer_devicedefault_callbackscbdefault_label_names num_devicess" r__init__zTrainer.__init__s & <&J KKKJ4 0 < `ASASA`A` /18~t0T\`0`1 el    2Ef - (GE;,Gb  (u0ENs$  *2XJ5K]K]KqKqK~K~JDDIDVDVDjDjDwDwCxz  $(#3#3E#: tyy>A (( p##E*t/A/A\E]E]/] !QRR&*%?%?"  " "(($$(;(;T]]''##).D & ++.EG_-`a $$4 5'  /<.G]M]*( 0  & &E#8$?CUCdCdd  & &udkk :  ! ! DII # **77>"/2  # # //199  !**=9DD 5/ 0-2-F-FD *-01AOAVAVAX1.D *$(#;#; .-J*,6))(@%  ( ( 49Stu u  !t~~'ATEVEVEbj  " #(B..0 $||   $~~::  {8,-1'28'>A $"7"7!"; KKb c  $Z -F4>>]^K^d   %=%++*:*:*J*JK$$st t"&   # $yy !rss(99 2 22NNB399==CUCUBVWDDHII;O&&)iimm&8&8%9; !$ 2 2DI399==&1NNB399==CUCUBVWcc II(C(Cv(M{{ell51199=()]^^2;D/ KK&!"&"<"<">!22<< ~M Q[\^`oQp   !% $** %$))*?*?*G NN;DJJ*>?26))2G2G2O.UYU^U^UjUj.tzz/C/CD,,88DJJPTP\P\] "&!6!6%*" 446   &@&BTU U&*&6&6&:&:=%&P#  & &) !MNN88:K   rwwrxxk0B'CkSTEUbtu v&*&>&>&btGbGbCb#e pX s A@1 A@1;A@6 A@6 returncDtjd|jS)NzUTrainer.tokenizer is now deprecated. You should use Trainer.processing_class instead.rr<rrs rrzTrainer.tokenizer.sno$$$rc<tjd||_y)NzjTrainer.tokenizer is now deprecated. You should use `Trainer.processing_class = processing_class` instead.r)rrs rrzTrainer.tokenizer3s x !1rc|jj|}t|r%|jjj }n|j }~|j |_|jt}||_ |S)z Activates the neftune as presented in this code: https://github.com/neelsjain/NEFTune and paper: https://arxiv.org/abs/2310.05914 ) rOr)r base_modelrget_input_embeddingsrSregister_forward_hookr`neftune_hook_handle)rrr embeddings hook_handles r_activate_neftunezTrainer._activate_neftune:sy **77> / *(3399NNPJ(==?J )-)A)A & 667PQ #.  rc"t|ds td|jj|}t |r%|j j j}n|j}|jj|` ~y)z^ Deactivates the neftune method. Make sure to call `_activate_neftune` first. rzNNeftune is not activated make sure to call `trainer._activate_neftune()` firstN) r;rrOr)rrrrrremoverS)rrrrs r_deactivate_neftunezTrainer._deactivate_neftuneMs{t23mn n**77> / *(3399NNPJ(==?J   '')  *Orc:|jj|y)ag Add a callback to the current list of [`~transformers.TrainerCallback`]. Args: callback (`type` or [`~transformers.TrainerCallback]`): A [`~transformers.TrainerCallback`] class or an instance of a [`~transformers.TrainerCallback`]. In the first case, will instantiate a member of that class. N)rYrZrcallbacks rrZzTrainer.add_callback^s **84rc8|jj|S)aK Remove a callback from the current list of [`~transformers.TrainerCallback`] and returns it. If the callback is not found, returns `None` (and no error is raised). Args: callback (`type` or [`~transformers.TrainerCallback]`): A [`~transformers.TrainerCallback`] class or an instance of a [`~transformers.TrainerCallback`]. In the first case, will pop the first member of that class found in the list of callbacks. Returns: [`~transformers.TrainerCallback`]: The callback removed, if found. )rY pop_callbackrs rrzTrainer.pop_callbackis$$11(;;rc:|jj|y)a Remove a callback from the current list of [`~transformers.TrainerCallback`]. Args: callback (`type` or [`~transformers.TrainerCallback]`): A [`~transformers.TrainerCallback`] class or an instance of a [`~transformers.TrainerCallback`]. In the first case, will remove the first member of that class found in the list of callbacks. N)rYremove_callbackrs rrzTrainer.remove_callbackys --h7rc|j|}|jjtjk(rt |dr|j yyy)N tie_weights)torrDrfTPUr;r)rrr8s rrLzTrainer._move_model_to_devicesF  99 " "l&6&6 675-;X    Y  KK()9(:;JJ((1122STXT]T]^mTnSopyy122HI]I]IfIfHgh77 0M18L8L3L1MM w<1 ==AYY=WB>S>S>_%%77:ei ( **TYY-R-RR**!1  !!3!34 4rc|j td|j}|j}tr.t |t j r|j|d}n|j|d}|j||jj|jj|jjd}t |tjj j"sN|j%|d<|jj&|d<t(|d<|jj*|d<|j,j/t1|fi|S) a@ Returns the training [`~torch.utils.data.DataLoader`]. Will use no sampler if `train_dataset` does not implement `__len__`, a random sampler (adapted to distributed training if necessary) otherwise. Subclass and override this method if you want to inject some custom behavior. z+Trainer: training requires a train_dataset.trainingr batch_size collate_fn num_workers pin_memorypersistent_workerssampler drop_lastworker_init_fnprefetch_factor)rrrryrrrrrr}rdataloader_num_workersdataloader_pin_memorydataloader_persistent_workersrrgrhrrdataloader_drop_lastrbdataloader_prefetch_factorrOpreparer)rrrdataloader_paramss rget_train_dataloaderzTrainer.get_train_dataloaders9    %JK K** ** "z-AQAQ'R 77 S]7^M CCM_iCjM00'99;;))99"&))"I"I  -)9)9)I)IJ+/+B+B+D i (-1YY-K-K k *2= . /37993W3W / 0'' =(VDU(VWWrcL| t|sy|jjrtr2t |t j t jStrGt |tjtj|jjSt|S|jjrtrXt!|t"j$r>|jj&|j(vr||jj&nd}nd}|j*|j*j,dnd}t/|jj0|||S|jj2dkr t|Sy)N) num_replicasrank)rrrrrr )r_ruse_legacy_prediction_looprr@xmxrt_world_size get_ordinalrrodp_sizedp_rankper_device_eval_batch_sizerriryrrrrrrrr?eval_batch_size world_size)rrrrs r_get_eval_samplerzTrainer._get_eval_sampler s\  z,'? 99 / /%'3 r/@/@/BIY)*3 !$#yyCC )66 99 $ $$&:lHDTDT+Uyy33|7P7PP!!=!=> >B>S>S>_%%77:ei ( ))$!1   99  1 $$\2 2rct||j tdt|tr|nd}t |drL||j vr>|j jr(|jj|j |St|tr|j|n||n |j}|j}tr.t|tjr|j|d}n|j|d}|j j ||j j"|j j$|j jd}t|t&j(j*j,sF|j/||d<|j j0|d<|j j2|d <t5|fi|}|j jr%t |dr||j |<n ||i|_|jj|S) a  Returns the evaluation [`~torch.utils.data.DataLoader`]. Subclass and override this method if you want to inject some custom behavior. Args: eval_dataset (`str` or `torch.utils.data.Dataset`, *optional*): If a `str`, will use `self.eval_dataset[eval_dataset]` as the evaluation dataset. If a `Dataset`, will override `self.eval_dataset` and must implement `__len__`. If it is a [`~datasets.Dataset`], columns not accepted by the `model.forward()` method are automatically removed. z-Trainer: evaluation requires an eval_dataset.eval_eval_dataloaders evaluationrrrrr)rrrstrr;rrrrOrrryrrrrrrrrrgrhrrrrr)rrdataloader_keyrreval_dataloaders rget_eval_dataloaderzTrainer.get_eval_dataloader6s  D$5$5$=LM M*4L#)FF D- .$"8"88 77##++D,B,B>,RS S,,   l +'"" ** "z,@P@P'Q66|Q]6^L CCM_kClM))33'99;;))99"&))"I"I  , (8(8(H(HI+/+A+A,+O i (-1YY-K-K k *37993W3W / 0%\G5FG 99 2 2t019H&&~6*8/)J&''88r test_datasetc|j}tr.t|tjr|j |d}n|j |d}|jj||jj|jj|jjd}t|tjjjsF|j!||d<|jj"|d<|jj$|d<|j&j)t+|fi|S)a Returns the test [`~torch.utils.data.DataLoader`]. Subclass and override this method if you want to inject some custom behavior. Args: test_dataset (`torch.utils.data.Dataset`, *optional*): The test dataset to use. If it is a [`~datasets.Dataset`], columns not accepted by the `model.forward()` method are automatically removed. It must implement `__len__`. testrrrrr)rryrrrrrrrrrrrrgrhrrrrrOrr)rrrrs rget_test_dataloaderzTrainer.get_test_dataloaderss** "z,@P@P'Q66|QW6XL CCM_eCfM))33'99;;))99"&))"I"I  , (8(8(H(HI+/+A+A,+O i (-1YY-K-K k *37993W3W / 0'' <(UCT(UVVrnum_training_stepsc|jtr;tjjj r|j j }n |j }|j||y)aZ Setup the optimizer and the learning rate scheduler. We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer's init through `optimizers`, or subclass and override this method (or `create_optimizer` and/or `create_scheduler`) in a subclass. )r rTN)create_optimizerrnrorprqr rTcreate_schedulerrr rTs rcreate_optimizer_and_schedulerz&Trainer.create_optimizer_and_schedulersO  $););00II 1CyYrc,t|tgd}|S)a0 Get all parameter names that weight decay will be applied to. This function filters out parameters in two ways: 1. By layer type (instances of layers specified in ALL_LAYERNORM_LAYERS) 2. By parameter name patterns (containing 'bias', 'layernorm', or 'rmsnorm') )bias layernormrmsnorm)rFr/)rrdecay_parameterss rget_decay_parameter_namesz!Trainer.get_decay_parameter_namess/u6JLlmrc tr |jn |j}|jE|j |}|j Dcgc]\}}||vs |j s|c}}|jjd|j Dcgc]\}}||vs |j s|c}}ddg}|j|j\}}n|j|j|\}}d|vr|jd}d|vr|jd}d|vr|jd}||fi||_|jdk(rddl }|jjj!} d} |j#D]} t%| t&j(s| t+| j-Dcic]!}|j/|j1#c}j3z } t4j7d | d | d z d | j9| d ddit4j;d| dt4j7d| d z d tr$t=j>|j|_|jScc}}wcc}}wcc}w)a Setup the optimizer. We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the Trainer's init through `optimizers`, or subclass and override this method in a subclass. N)r weight_decayrroptimizer_dictAdam8bitrzskipped : izM paramsweight optim_bits zbitsandbytes: will optimize z in fp32z skipped: ) rrNrrTrnamed_parameters requires_gradrrrget_optimizer_cls_and_kwargspopr2 bitsandbytesoptimGlobalOptimManager get_instancemodulesrr Embeddingsumrdata_ptrnumelr6rrregister_module_overridedebugroDistributedOptimizer) r opt_modelrnpoptimizer_grouped_parameters optimizer_clsoptimizer_kwargsr%managerskippedmodules rrzTrainer.create_optimizers+B*CD&& >> !#==iH '0&@&@&B"aqL\G\abapap%)II$:$: '0&@&@&B"aqP`G`efetet%(  , (,,8262O2O/ /262S2STXT]T]_h2i/ /++/?/C/CH/M,**/?/C/CG/L, #33/?/C/CDT/U,*+G\K[\DN%%3#&,,??LLN'//1VF!&",,73IZIZI\']A aggi(?']'d'd'f#gg hvhb58I$RS88LZ\K]^ 'CF88%TU V  i%'8AB " $ 55dnnEDN~~i J(^s* J/! J/.J/ J5, J59J5 &J; cVtd|jjDS)z9 Get the number of trainable parameters. c3VK|]!}|js|j#ywr)r"r-)rr3s rrz7Trainer.get_num_trainable_parameters..sQ1779Qs)))r+rrrs rget_num_trainable_parametersz$Trainer.get_num_trainable_parameterss"Qdjj&;&;&=QQQrc|j td|jjDcgc]}|d c}Scc}w)zR Returns the learning rate of each parameter from self.optimizer. PTrainer optimizer is None, please make sure you have setup the optimizer before.lrrTrrV)rgroups rget_learning_rateszTrainer.get_learning_ratess< >> !op p)-)D)DEd EEEs ?rc|j td|&|jjD] }||dvs |cS|jjDcgc]}|d c}Scc}w)a Returns optimizer group for a parameter if given, else returns all optimizer groups for params. Args: param (`str` or `torch.nn.parameter.Parameter`, *optional*): The parameter for which optimizer group needs to be returned. r>rr@)rrrAs rget_optimizer_groupzTrainer.get_optimizer_groupsn >> !op p  44 !E(O+ L !.2^^-H-HIEhIIIs A'c#$i#jrIjjddjdD]}|jd\}}|#|<dji$jj fj d} d{dtd tttfd tttfd td tttff #$fd }jtjk(rt}$j!ddd|$fSjtj"tj$fvrMddlm}|}$j!|jtj$k(r$j!ddi|$fSjtj*k(r ddlm}|}$j!||$fSjtj2k(r ddlm} | }$j!||$fSjtj8k(r ddlm} | }$j!||$fSjtj>tj@tjBtjDtjFtjHtjJtjLtjNtjPtjRtjTtjVtjXtjZfvrk ddl.m}m/} m0} d} d}d}|}djvrd} djvrd}djvr|}ndjvr| }d jj fi}nVd!jvr| }#}nBd"jvr3tcrStejfthjjjed#tejfd$kr t1d%dd&l.m6}|}to#jqd'jto#jqd(j to#jqd)d*fto#jqd+d,to#jqd-j d.}d/#vrts#d/|d/<d0#vrts#d0|d0<d1|i}d!jvr| |d2<$j!|$j!|tcr]tejfthjjjed#tejfd4krttjwd5|$fSjtjxk(r dd6l=m>}|}$j!|$j!t#jqd7d8tt#jqd9d:tt#jqd;d:tt#jqd|$fSjtjk(rtjj}|$fSjtjk(rtjj}|$fSjtjk(rtjj`}|$fSjtjtjtjtjtjtjfvr/ts t/d@ddAlMmN}mO}mP}tj|tj|tj|tj|tj|tj|i}ts#jdBdCts#jdDdEto#jdFdG#jdHdIdJ}|j||\}$jtjk(r$j!ddd|$fSjtjtjfvrts t/dKddLlUmV}tj|tj|i}ts#jdBdC#jdMdN#jdOdPts#jdDdEto#jdFdQ#jdHdIdR}|j||\}$|$fSjtjtjfvrjts t/dStdTs t/dU t1dVddWl[m\}m]}dXjvr|}n|}$j!dYi|$fSjtjk(rts t1dZdd[l`ma}|}$j!to#jqd\d]to#jqd^d_to#jqd`dato#jqdbdato#jqdcdQdd|$fSjtjtjfvrtrHtejfthjjjedetejfdfkr t/dgtejfthjjjedhtejfdikr t/djddklemf}mg}jtjk(r|}n+jtjk(r|}n t1dl$j!||$fSjtjtjtjfvr?ts t/dmtdTs t/dnddollmm}mn} i}d}!jtjk(r#tdps t/dqddrllmo}"|"}|}d}!nMjtjk(r|}|}n+jtjk(r| }n t1dsj|dt<|!rj|du<|j!to#jqdvd_to#jqdwdxdy$j!||$fSt1dzj#t.$r t1dwxYw#t.$r t1dwxYw#t.$r t1dwxYw#t.$r t1d3wxYw#t.$r t1d?wxYw)|z Returns the optimizer class and optimizer parameters based on the training arguments. Args: args (`transformers.training_args.TrainingArguments`): The training arguments for the training session.  r,=r?)betasepsToptimizer_nameoptimizer_mapping optim_kwargsis_layerwise_supportedrc|jjd}|r.jtjk(r|rt d|d||}j td|dtj ttfstdj td|dtj txrj jd d d k(}g}g}jD]\} } tj | d \} } t| tjs | r| st j#| d|d[| s|s`|j%| j&|j%| dzt)|dk(rtd|dj dj+D cgc] \} }| |vs |}} }|j-d|id|i|g}|rj.dk7rtd|di|D]}|d|gigfi|<|D]}|d|gi|gfi|<fd}j1D] }|j2s|j5|"t6}j-dij-d|i|fScc}} w)a Helper function to set up low-rank optimizers like GaLore and Apollo. Args: optimizer_name (str): Name of the optimizer. optimizer_mapping (dict): Mapping of optimizer names to their classes. optim_kwargs (dict): Keyword arguments for the optimizer. is_layerwise_supported (bool): Whether layerwise optimization is supported. Returns: Tuple[Any, Any]: Optimizer class and updated optimizer kwargs. layerwisez Layer-wise z" does not support DDP at this timez1You need to define `optim_target_modules` to use z optimizerszX`optim_target_modules` must be a list of strings, a regex string, or 'all-linear'. Got: z'You need to pass a model to initialize z optimizer._-z all-linearT)return_is_regexz matched but ignored. z only supports linear layers.z.weightrzNo target modules found for z ().rr z Layerwise z( does not support gradient accumulation!cl|j'|j|jyyr)gradstep zero_grad)rrs roptimizer_hookz^Trainer.get_optimizer_cls_and_kwargs..setup_low_rank_optimizer..optimizer_hook}s4zz-&u-224&u-779.rr)lowerendswithrDrfrENotImplementedErroroptim_target_modulesrrrrreplace named_modulesrYrLinearrr<appendrr7r!updaterrr""register_post_accumulate_grad_hookr>)rKrLrMrN is_layerwiser5 all_linear target_paramstarget_params_names module_namer9target_module_existsis_regexr2r3non_target_paramsrVrrYrrr optim_argsr6s @rsetup_low_rank_optimizerzFTrainer.get_optimizer_cls_and_kwargs..setup_low_rank_optimizer,sX$*//1::;GL 2 2l6N6N NSi)K7GGi*jkk-n=M((0 #TUcTddo!pqqd77$E nospIpIoJK} #J>JZZe!fgg444c:P--55c3?<O  M"$ ',':':'< D# V1K--{D2.$h"&"))4+H*m+A.AQQno+J$$V]]3#**;+BC! D$=!Q& #??OrRVRkRkQlln!opp/4/E/E/G htq!1TgKg h  h    +,-=9L9L 33q8$z.1AAi%jkk!#.eE,9Hug;N:O,dSc,dN5)e*uE,9Hug;^Q];^:_,tcs,tN5)u: #--/QE**@@PQ!8  '')9>(JK  # #X|$< = "22 2?!is  K%KF)scale_parameter relative_stepr)AdamWfusedz7Trainer failed to import syncfree AdamW from torch_xla.) NpuFusedAdamWz3Trainer failed to import FusedAdamW from torch_npu.) FusedAdamzFTrainer tried to instantiate apex FusedAdam but apex is not installed!)rpLionRMSpropr Npaged8bitadamlionrIrmspropademamixr%z0.44.0z{The AdEMAMix optimizer is not supported by your current version of `bitsandbytes`. Please install `bitsandbytes` >= 0.44.0.)AdEMAMixbeta1beta2beta3gH.?alphag@rJ)rIrrJt_alphat_beta3ris_pagedzOTrainer tried to instantiate bnb optimizer but `bitsandbytes` is not installed!z0.41.1zYou are using 8-bit optimizers with a version of `bitsandbytes` < 0.41.1. It is recommended to update your version as a major bug has been fixed in 8-bit optimizers.)AnyPrecisionAdamWuse_kahan_summationFalsemomentum_dtypefloat32variance_dtypecompensation_buffer_dtypers)rrrrz4Please install https://github.com/pytorch/torchdistxzYou need to install `galore_torch` in order to use GaLore optimizers install it with `pip install git+https://github.com/jiaweizzhao/GaLore`)GaLoreAdafactor GaLoreAdamWGaLoreAdamW8bitrupdate_proj_gapscaleg? proj_typestd)rrrrzYou need to install `apollo_torch` in order to use APOLLO optimizers install it with `pip install git+https://github.com/zhuhanqing/APOLLO`) APOLLOAdamWprojrandom scale_typechannelg?)rrrrrrziYou need to install `lomo_optim` in order to use LOMO optimizers install it with `pip install lomo-optim`0.30.0zGYou need to have `accelerate>=0.30.0` to be able to use LOMO optimizerszMYou need to pass a `model` in order to correctly initialize a LOMO optimizer.)AdaLomoLomoadarz5Please install grokadamw with `pip install grokadamw`) GrokAdamW alpha_initg\(\?lambg@gammag?grokking_signal_decay_rategradient_clipping)rrrrrtorchaoz0.4.0zYou need to have `torchao>=0.4.0` in order to use torch 4-bit optimizers.Install it with `pip install torchao` or follow the instructions here: https://github.com/pytorch/aorz2.4zYou need to have `torch>2.4` in order to use torch 4-bit optimizers. Install it with `pip install --upgrade torch` it is available on pipy. Otherwise, you need to install torch nightly.) AdamW4bit AdamW8bitzInvalid optimizerzwYou need to install `schedulefree` in order to use schedulefree optimizers. Install it with `pip install schedulefree.`zOYou need to have `accelerate>=0.30.0` to be able to use schedulefree optimizers)AdamWScheduleFreeSGDScheduleFreerzYou need to install `schedulefree>=1.4.0` in order to use RAdamScheduleFree optimizer. Install it with `pip install schedulefree.`)RAdamScheduleFreezInvalid schedulefree optimizerr warmup_stepsweight_lr_powerrr)rrz2Trainer cannot instantiate unsupported optimizer: T)rrlr^split learning_rate adam_beta1 adam_beta2 adam_epsilonrdictrbooltupler&re ADAFACTORr,rb ADAMW_TORCHADAMW_TORCH_FUSED torch.optimrpADAMW_TORCH_XLAtorch_xla.amp.syncfreer=rADAMW_TORCH_NPU_FUSEDtorch_npu.optimrrADAMW_APEX_FUSEDapex.optimizersrs ADAMW_BNB ADAMW_8BIT PAGED_ADAMWPAGED_ADAMW_8BITADEMAMIX ADEMAMIX_8BITPAGED_ADEMAMIXPAGED_ADEMAMIX_8BITLION LION_8BIT PAGED_LIONPAGED_LION_8BIT RMSPROP_BNB RMSPROP_8BIT RMSPROP_32BITbitsandbytes.optimrtrurxrrrrr}floatrintrr<ADAMW_ANYPRECISIONtorchdistx.optimizersrrr3rSGDADAGRADAdagradRMSPROP GALORE_ADAMWGALORE_ADAMW_8BITGALORE_ADAFACTORGALORE_ADAMW_LAYERWISEGALORE_ADAMW_8BIT_LAYERWISEGALORE_ADAFACTOR_LAYERWISErz galore_torchrrrr$ APOLLO_ADAMWAPOLLO_ADAMW_LAYERWISErw apollo_torchrLOMOADALOMOrru lomo_optimrr GROKADAMWr{ grokadamwrADAMW_TORCH_4BITADAMW_TORCH_8BITrtorchao.prototype.low_bit_optimrrSCHEDULE_FREE_RADAMSCHEDULE_FREE_ADAMWSCHEDULE_FREE_SGDr schedulefreerrrrr)%rrmappingkeyvalue adam_kwargsrmr5rprrrsrtrurradditional_optim_kwargsr} bnb_kwargsrrrrrLgalore_optim_kwargsrapollo_optim_kwargsrrrrrrrrequire_warmuprrlr6s%`` @@rr#z$Trainer.get_optimizer_cls_and_kwargssh  ????223;AA#F ($]]3/ U"' 3 (!$"4"45oot7$$ ,0 ^ 3^ 3#CH~^ 3sCx.^ 3%) ^ 3 38_ ^ 3^ 3@ ::11 1%M  # #PU$V Wp ...o ZZN668X8XY Y )!M  # #K 0zz^=== ''$8b ...a ZZ>99 9 \8 %  '' 4V ...Q ZZ>?? ? X9 -  '' 4F ...A ZZ>:: : k5 )  '' 4v...qZZ  $ $  % %  & &  + +  # #  ( (  ) )  . .     $ $  % %  * *  & &  ' '  ( (  "8 tCC  $ *5'djj(#HTZZ'!"JTZZ'$)Mtzz)$(M/6$//8Z.[+$**,$+M/9+4::-02w}}!**22>B8 h/80)G <$,M"*..$//"JK!*..$//"JK!*..&"AB" "'z~~gs'C!D$Z^^E4;L;L%MN/+!J.=@IAV=W/ : J.=@IAV=W/ :*J7 DJJ.-5Jz* ''(?@ '' 3)*w}}""**>:0 h'0(rV...OZZ><< < YC 1  '' 4!''/8H]_f9g/h*1%HXZc9d*e*1%HXZc9d*e5<!:>>2Mz#Z6  ~...gZZ>-- -!KKOOMd...cZZ>11 1!KK//M`..._ZZ>11 1!KK//M\...[ZZ  ' '  , ,  + +  1 1  6 6  5 5   -.!_ S R++[00///55{::O99? ! JNN6378#&z~~6G'M#Nz~~gt<='^^K? # /G -/B/ +M+zz^<<< ''ETY(Z[T...SZZ  ' '  1 1  -.!^ 1++[55{!  JNN6378"vx8(nn\9E#&z~~6G'M#Nz~~gs;<'^^K? # /G -/B/ +M+`...[ZZN//1G1GH H$&!@+84!"kll} !pqq 0 " ' $  # #We$4 5t...sZZ>33 3)+ !XYY +%M  # #"' |T(J"K!*.."=>":>>'3#?@27 Gceh8i2j).z~~>QSV/W)X  d...SZZ  + +  + +  ()W]]9;M;M;U;UV_;`-adkdqdqe."{}}Y//77@AW]]SXEYY!K Mzz^<<< ) ~>>> )  !455  # #K 0b...aZZ  . .  . .  , ,  -.!B+84!"stt G&( #!Nzz^???09%F; 1 *5'!&~AAA 1 *5'~??? /  !ABB6:6G6G #N 3:>:K:K'7 # * *',Z^^1w4,x 'x$5G?x<=B"y4x  x!$x9<yy)rTc|jkt|jj| |jn||jj |||jj |_d|_|jS)z Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or passed as an argument. Args: num_training_steps (int): The number of training steps to do. )rTnum_warmup_stepsr scheduler_specific_kwargsT)rUr-rlr_scheduler_typerTget_warmup_stepslr_scheduler_kwargsr~rs rrzTrainer.create_schedulersq    $ - ++,5,=$..9!%!;!;95 Kz?TYY%J%JJ J Ks:AA6B  B train_dlrecd} |D]#}|dj}|||zcS||z }% |S#t$rtjdY|SwxYw)zq Helper to get number of tokens in a [`~torch.utils.data.DataLoader`] by enumerating dataloader. r input_idsz%Cannot get num_tokens from dataloader)r-KeyErrorrr<)rre train_tokensbatchtokenss r num_tokenszTrainer.num_tokenssz   D! '{+113(!I--&  ' D NNB C Ds //AAtrialz optuna.Trialc ||_|j|y|jtjk(r|j |}n|jtj k(r|}|j ddn|jtjk(rI|jjDcic]#\}}|t|tr t|n|%}}}n|jtjk(r|}jD]v\}}t|j|st j#d|d6t%|j|d}|t'||}t)|j||x|jtjk(r"t j+d|j,|jtjk(r"t j+d|j|jtjk(rt j+d||j.r|jj0 t3d|j4j7d d lm}d d lm} | |jj0|j_ |jj@jC|j||jj@ |j_"tGjI|jKycc}}w) zHP search setup codeNwandbzTrying to set zY in the hyperparameter search but there is no corresponding field in `TrainingArguments`.zTrial: zSigOpt Assignments: zW&B Sweep parameters: z7For sweeps with deepspeed, `args.deepspeed` must be setr)DeepSpeedPluginHfTrainerDeepSpeedConfig) hf_ds_config)&_trialryrROPTUNAhp_spaceRAYr$SIGOPT assignmentsitemsrrrWANDBr;rrr<r3rsetattrrrrCr#rrO free_memoryaccelerate.utilsr#transformers.integrations.deepspeedrhf_deepspeed_configtrainer_config_processdeepspeed_pluginr _reset_stater%) rr rrvrrold_attrrrs r_hp_search_setupzTrainer._hp_search_setups  ! ! )U]   ! !_%;%; ;]]5)F  # #':': :F JJw %  # #'='= =INIZIZI`I`IbcAa:a#5Q1<cFc  # #'<'< <F ,,. +JC499c*$SE*,,tyy#t4H#&Xu- DIIsE * +  ! !_%;%; ; KK'%,,0 1  ! !_%;%; ; KK.u/@/@.AB C  ! !_%:%: : KK08 9  $ $yy""* !Z[[    ( ( * 9 T,DTYYEXEX,YDII ) II ) ) @ @ K)8diiFcFc)dDII &   + + - //1Uds2(L;rWmetricsc~|j|y|j}|j||_|jtj k(rddl}t|dr|jjsy|j|j||jrL|jj|j|j|j |j"yyy|jtj$k(rddl}t)j*5}d}|j j,r7|j/||j0j2j5|}|j|d<|j0j||dddyy#1swYyxYw)Nrstudy)checkpoint_dir objective) checkpoint)rycopycompute_objectiver*rRroptunar;r(_is_multi_objectivereport should_prunerY on_train_endrrprw TrialPrunedr ray.traintempfileTemporaryDirectoryra_tune_save_checkpointtrain Checkpointfrom_directory)rr rWr&r.raytemp_checkpoint_dirr+s r_report_to_hp_searchzTrainer._report_to_hp_search1ss  ! ! )U] ,,.//8  ! !_%;%; ; ug&u{{/N/N/P T^^T2%%'))66tyy$**dll[,&,,..(0Q&  # #':': : ,,. A2E! <<++..>Q.R!$!5!5!D!DEX!YJ'+~~ $   Z @  A A; A As -ABll>P>P>RDJJ ) )*: ; JJ # #BGGLL=O$P Q JJt~~002BGGLL^4\ ] JJt((335rww||JP^7_ ` !rct|j}|dk(r|j}n"|dk(r|j|}n td| td|S)Nrr z'model_init should have 0 or 1 argument.z"model_init should not return None.)rarr-)rr model_init_argcountrs rr,zTrainer.call_model_initSs]1$//B ! #OO%E A %OOE*EHI I =CD D rFc ~|sK|tjd|Stt|}|j |} t j |}|j |jjdd}|r||_ td}|jj|5tj5tj tj tj"j$tj dk\rht'|t(r#tj*j-||d}ntj*j-||Dcic]}||| c}d}n_g} |D]+}tj.||} | j1| -t3| } tj*j-|| d}ddddddtj*j5|}tj5|d i||d i|ddd|}d|_|S|Scc}w#1swYoxYw#1swYsxYw#1swY3xYw#t8t:t<t>t@f$r$} tjd | d Yd} ~ |Sd} ~ wwxYw) NzAfailed to use PyTorch jit mode due to current dataloader is none._original_forwardF) cache_enabled)autocast_handlerr)example_kwarg_inputsstrictrQz'failed to use PyTorch jit mode due to: .r)!rr<nextiter_prepare_inputsr,r__dict__r$rPrrOautocastrno_gradrrr base_versionrrjittrace ones_likerarfreezerlr-rrr IndexError) rrrr example_batch jit_modeloriginal_forwardrOr jit_inputsexample_tensores rtorch_jit_model_evalzTrainer.torch_jit_model_evalaso!bc  j!12M 00?M O IIe,  #,#5#5#9#9:Mt#T #(8I%#1#F %%..@P.QYSXS`S`SbY}}W]]53D3D%E%R%RSW^WdWdelWmm%mT:(-  Xens(tI(-  )Yf5gRUc=;M6M5g',)8)I &( #0>C-2__]3=O-PN&--n=>&+:%6 $)IIOOIzRWO$X !YY""II,,Y7 ]]_/. .. ./"#(  u '6hYYYY$// !)ZJO O!H1MNN  OsuA2I:5I" B(I2 I?A)I(I"0;I:+I.<I:II I""I+'I:.I73I::J<J77J<cts tdddl}|sg|j|js&|j j rtjn|}|j||dd|j }|S|js|j|j|||jdd\}|_ |S)NzUsing IPEX but IPEX is not installed or IPEX's version does not match current PyTorch, please refer to https://github.com/intel/intel-extension-for-pytorch.rO1F)rlevelconv_bn_foldinginplaceT)rrTrkri) r}r=intel_extension_for_pytorchrr$rrHrrsoptimizerr8rT)rrrripexs ripex_optimize_modelzTrainer.ipex_optimize_models "L  3 JJL*.*:*:tyy?W?WENN]bEMM%uDRWeieueuauMvE >> $(MMUdnndRV%2% !E4> rc jdddd}d}d}|jD]<\}}t||d}t||d} |#| &|| k7s,|d|d|d | d z }d }>|j} |jt d |j z} | | k7r|d | d | d z }d }|rt j|yy)N logging_steps eval_steps save_steps)rqrrrsFztWarning: The following arguments do not match the ones in the `trainer_state.json` within the checkpoint directory: z rz (from args) != z (from trainer_state.json)Tr z per_device_train_batch_size: )rr3rr|maxn_gpur warning_once) r training_args trainer_stateattributes_map has_warning warning_strarg_attr state_attr arg_value state_value train_bs_argstrain_bs_states r#compare_trainer_and_checkpoint_argsz+Trainer.compare_trainer_and_checkpoint_argss,&&   M $2$8$8$: # Hj x>I!-TBK$)@YR]E]hZr)}oM]^l]mnHI IKK     , rc jjr@jrtjntj }j |||}tretjtjjr jStj|jjSjj||ur|Sj r@|r>t#j$|j&jj(\}_jj*dkDr"t-|ddst/j0|}jj2rMt5j4}j7|||}t9t5j4|z d_|s|Sj<r dd lm dd lm!dd l"m#}m$}jJrdd l&m'd}d} t-|dd} jjRjUd| } jjRddkDr/tWjX|jjRd}nT| Rt[} | D],} t]|| }| t_d| ja|.tWjX|| }jjb}jjRdrD|jdjfr&thjkdd|jd_3fd} jJrd}|||| x_ }n|f|| d|x_ }difd}|tl_7|Stqr@t.jrju|twtyjzdg}|Sjj|t~jk(rtr|Si}jjjj|d<n&t|tr|j |d<nd|d<jjjj|d<jjjj|d <td!i|j_H|S#tP$r tQd wxYw)"N)r)backward_passes_per_step) opt_levelr is_loaded_in_8bitFr)XlaFullyShardedDataParallel)checkpoint_module)size_based_auto_wrap_policytransformer_auto_wrap_policy)SpmdFullyShardedDataParallelzJMissing XLA FSDP related module; please make sure to use torch-xla >= 2.0._no_split_modulestransformer_layer_cls_to_wrapmin_num_params)rz@Could not find the transformer layer class to wrap in the model.)transformer_layer_clsxla_fsdp_grad_ckptzX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.cHjsn}||g|i|Sr)r)mrkwargs target_clsFSDPFSDPv2rrs rauto_wrapper_callablez2Trainer._wrap_model..auto_wrapper_callables--1-H-HfJ%&7&:LTLVLLrcddlm}d}t|tjr|}n.t|t r|d}nt||r |j }| tdtj||dy)Nr )CausalLMOutputWithPastrzASomething went wrong, the output of the model shouldn't be `None`)rNN) modeling_outputsrrrTensorrlogitsrr mark_sharding)outputmeshr real_outputs r shard_outputz)Trainer._wrap_model..shard_outputsiH"&K!&%,,7&, #FE2&,Qi #F,BC&,mm "*()lmm$$[$8LMr)rauto_wrap_policyr)rrcV|jdi|}|rtj|S)Nr)rWr mark_step)rTbarrieroptimizer_argslosss rpatched_optimizer_stepz3Trainer._wrap_model..patched_optimizer_step9s&%y~~77LLN rSMDATAPARALLEL_LOCAL_RANK) device_idsfind_unused_parametersT bucket_cap_mbbroadcast_buffersr)Iruse_ipexrlrrsrrorrrNrorDistributedModelrrOr)rkr initializerTfp16_opt_levelrur3r DataParallel jit_mode_evaltimerfroundjit_compilation_timerBtorch_xla.distributed.fsdprrtorch_xla.distributed.fsdp.wraprrr7torch_xla.experimental.spmd_fully_sharded_data_parallelrr=rAr functoolspartialr5rE Exceptionaddxla_fsdp_configconfig use_cacherrvroptimizer_steprparallelDistributedDataParallelrrbgetenvrDrfrErddp_find_unused_parametersr'is_gradient_checkpointingddp_bucket_cap_mbddp_broadcast_buffersr ddp_handler)rrrrr start_timerrrr%default_transformer_cls_names_to_wrap"fsdp_transformer_layer_cls_to_wraptransformer_cls_to_wrap layer_classtransformer_cls fsdp_kwargsrrrrrrs` @@@r _wrap_modelzTrainer._wrap_models{ 99  &*&6&6ENNEMME,,UHE,JE " $$,,cii.H.HI)))'' HmHmn n    ( ( /u <L ==X$'NN5$..TXT]T]TlTl$m !E4> 99??Q wu6I5'QOOE*E 99 " "J--eZJE(-diikJ.F(JD %L  # # pZH .. $ $( !4;ECVX\4] 1151F1F1J1J/1V2 .yy$$%56:#,#4#4/ @U@UVf@g$ 4?*-%'#EEK&@ &TO&.'(jkk/33OD E$-#4#40*A$ ))33Kyy$$%9:<<))''r.3ELL*M ** N&,!-%5*? & U&*&%5*?&" & U;@PR  !7B 4 3% &KK773ryy1L'M#N"O8E0 +YY $ $ (@(@ @,. Fyy33?37993W3W/0E?38=7V7V3V/037/0yy**6*.))*E*E'yy..:.2ii.M.M*++H+R6+RD   ( U p!"noo ps %&S,,Tresume_from_checkpointignore_keys_for_evalc |durd}|jj|j}d|_|j |j |j |_|js |jrJ|js>|js2|j&|j|j |jd|vr+|jd}tj dt"t%|dkDr5t'ddj)t+|j-d |j/||jj0|_d}|j{|jj4rt7|jj8nt;|jj8|j=||_d}d \|_|_ tC|tDr2|r0tG|jH}|tKd |jHd |tMs)|jNs|jPs|jS|tUjVtXjZj)|t\}|j0|j0|_|rC|j^r&|j|j |j|j |_0tc|jd|j2|jf}|jhr5 tkjl||||| tkjnS||||| S#tkjnwxYw)a Main training entry point. Args: resume_from_checkpoint (`str` or `bool`, *optional*): If a `str`, local path to a saved checkpoint as saved by a previous instance of [`Trainer`]. If a `bool` and equals `True`, load the last checkpoint in *args.output_dir* as saved by a previous instance of [`Trainer`]. If present, training will resume from the model/optimizer/scheduler states loaded here. trial (`optuna.Trial` or `Dict[str, Any]`, *optional*): The trial run or the hyperparameter dictionary for hyperparameter search. ignore_keys_for_eval (`List[str]`, *optional*) A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions for evaluation during the training. kwargs (`Dict[str, Any]`, *optional*): Additional keyword arguments used to hide deprecated arguments FNT model_pathzi`model_path` is deprecated and will be removed in a future version. Use `resume_from_checkpoint` instead.rz*train() got unexpected keyword arguments: rrSrz/No valid checkpoint found in output directory ())rrr r)8r'r(rr$rSrrrGrHrIr4rrLr8r$r.r/r0r7rrrrr%r|r}r r\r!rcr,rTrUrrr^rrrrCrJ_load_from_checkpointr9load_from_jsonrbrArErFrNr]_inner_training_loopauto_find_batch_sizer_ hf_hub_utilsdisable_progress_barsenable_progress_bars) rrr rrrmodel_reloadedrpinner_training_loops rr8z Trainer.train[s. "U *%) " ""$yy  # # /// ;DJ  D$7$7MM**'  & &tzz4;; ? 6 !%+ZZ %= " MM  v;?HSWX^XcXcXeSfIgHhhijk k e$!%!;!; ?? &7;yy7Q7Q #DIINN 3W_`d`i`i`n`nWo--e4DJ!N0: -DND- ,d 38N%8%I "%- #RSWSbSbRccd!eff ! -*,T5N5NW[WkWk**+AB // =SUg0hiE%%1).)?)?& ))**4::t{{C!%D 8  % %t'='=t?X?X     4224*+A)= 113&'=%9  113s 1M11Nc 0|jj||_|jjr|j j |jk7rddlm}||j\|_ |j|_ |jrt|jj}|jtd|jjz|j_ |jd||j_ |j|j _t j#d|j|j%}|j&r t)|}|j|j*z|j,z} |j/||| \} } } } }}}d}|jj0r9|j3||rdn|}||r||j4z}n||j*z}t6j8|jj"vr9|jjdkDr t;dt=|j}t?xs|j@xs |jB}|jBxr.tE|jj jFdddk(}|rd }|jHrd|_%d |_$|jrtM|| \|_'|_%|s|jQ| tS|jTjV|jXgzDcgc]}t[|t\s|c} |_|du|j _/|j|j _|j ja|||jbr&|jje|jf |ji|j}||jurdnd }|r(|jBrtk|jd |_ |rg|rS|jm|jjndk7r*|jjq|j|_ |jQ| |r|jjstu|jJdrk|jvr&|jjq|j}n|jjq|j|jN\}|_'n|jjq|j|jN|jJ\}|_'|_%n`|jjxtzj|tzj~fvr*|jjq|jN|_'|jBr|x|_ |_ ||jur||_ |jr|j|_@|k|jr-t|j|t|j n2t?s |jBr|j||j|j||j|t jdt jd| dt jd| dt jd|jjd|jj|jk7r#t jd|jdt jd| dt jd|j*t jd|dt jdt|ddd|j _Htj}d}d}d}|tjjtjj|trLtSjtjj|t|_|j|j|j |jt|j j| z}|js)|j j| z}||j*z}nd}t jdt jd|t jd|j j|jst jd|d |d!d"D]#}t|jT|tE||%||jT_V|j j||| |tjd#|j$}d#|_[|j j|_\|jd}d}|jTj||j |jX|_,|jr|j||d%t|| D ]} |}!tu|!d&r|!j| |jdk\rd|_d| t|!n|j|j*z}"|jTj||j |jX|_,| |k(r||dk(r|j|d }#d}$|dkDrt|!|}!|}$d}d}#d'}%t|!}&| |j*z}'|'dk(r |j*}'d'}(|"|j*zdz})|j*dk(r|)dz})t|)D]}*|(dz }(|(|)dz k7r |j*n|'}+|j|&|+|j\},}-t|,D]\}.}/|%dz }%|%dz|j*zdk(xs|%dz|"k(}0|jjj|0|jjrtE|jd(d)}1|1|/vrt jd*n|/|1j}2tj|2|jjtj+}2|j xj|jj|2jjz c_s|#r|j|d }#|dkDr0|dz}||jd|dk(r|j|||jd}|%|j*zdk(r6|jTj||j |jX|_,|.t|,dz k7rR|jjtjk7r+tj|jj|,ntj}3|35|j||/|-}4ddd|jrdt sYtj 4stj |4r-||d|j jz|jz z z}nC|j4jk7r%t;d-|jd.|4j||4z}|xjt|j|/z c_|0r|jjjd|jC|jdkDr2t?r5|jr(|jNj|j}5n|jvrNtjjt!j"|jN|j}5n7|jj|j%|j}5t'rU|jjtjk(r.|j)}tu|d/r|j}n|5}|jTj+||j |jX|_,|jNj-|jTj/||j |jX|_,|j1}|jj2sTt[|jJtjxjJj4s|jJj-|j|j xjdz c_S| |%dz|$z|"z z|j _H|jTj7||j |jX|_,|j9||||| |||0n7|jTj;||j |jX|_,|jXj<s|jXj>sސt rtAjBn|jXj<s|jXj>s|t rtAjBn|%dkrAt jd1|j jd2|d3d|jX_|jTjE||j |jX|_,|j9||||| |||0t6jF|jj"vrLt r,tAjHtKjLnt jd4|jXj>s n|jrtu|d5r tO|d5t jd6|jPr|j jRt rtAjTd7nW|jVtXjZk(rt]j^n t?rtaj^|jc|xj|jz c_[t|j jd8}6|j|6z }7ted9|| |j j|:}8|jg|j jh|8d;<|7|8d<<d |_|jljo|8|jq|8|js|}9|jud |9=}:|jjvr|j jR|jjxdk(rp|:D]k};tjj{|;|j jRr:t jd>|;d?t}j~|;d@m|jTj||j |jX|_,|j|j|j|jt|j j|7|8Scc}w#1swY KxYw)ANr)release_memoryr Tz)Currently training with a batch size of: zCurrently --debug underflow_overflow is not supported under DP. Please use DDP (torchrun or torch.distributed.launch (deprecated)). fsdp_versionF)r )r)gradient_checkpointing_kwargs) recursivefp8rWload_module_strictz***** Running training ***** Num examples = rGz Num Epochs = z( Instantaneous batch size per device = zA Training with DataParallel so batch size has been adjusted to: zE Total train batch size (w. parallel, distributed & accumulation) = z Gradient Accumulation steps = z Total optimization steps = z# Number of trainable parameters = )trainable_onlyzE Continuing training from checkpoint, will skip to saved global_stepz! Continuing training from epoch z' Continuing training from global step z Will skip the first z epochs then the first z batches in the first epoch.)rrTrUrr8)skip_scheduler set_epochmain_input_namerzTried to track the number of tokens seen, however the current model is not configured properly to know what item is the input. To fix this, add a `main_input_name` attribute to the model class you are using.)r8rrz0Calculated loss must be on the original device: z but device in use is item)rzXThere seems not to be a single sample in your epoch_iterator, stopping training at step zI! This is expected if you're using an IterableDataset and set num_steps (z.) higher than the number of available samples.zYou enabled PyTorch/XLA debug metrics but you don't have a TPU configured. Check your training configuration if this is unexpected._pastzU Training completed. Do not forget to share your model on huggingface.co/models =) rgMbP?r8) num_samples num_stepsr  total_flos train_loss use_mtimerDeleting older checkpoint [] due to args.save_total_limit ignore_errors)rOrr}rrrpr|rrrNrrCrrtrupropagate_args_to_deepspeedrr/rrr%rrset_initial_training_valuesinclude_tokens_per_secondr rfrUNDERFLOW_OVERFLOWrrrrBrJr3 fsdp_pluginr~rUr"rTrr9rYrrwrr4is_hyper_param_search compute_stepsgradient_checkpointinggradient_checkpointing_enablerrr)_fsdp_qlora_plugin_updatesmixed_precisionrr8r;rkr&rerrr#r#rr_load_optimizer_and_scheduler _load_scalerrrDepochrrbrAisfilerrErr_load_callback_staterrBignore_data_skiprtrain_dataloaderinit_training_referencesrrr8_total_loss_scalar_globalstep_last_loggedrXon_train_begin eval_on_start _evaluaterr past_indexrr7reon_epoch_begin_load_rng_staterrUget_batch_samples enumerategradient_state_set_sync_gradientsinclude_num_input_tokens_seenr<r-int64num_input_tokens_seengatherr+rrbclose on_step_begindistributed_typer DEEPSPEEDrrno_syncrr training_steplogging_nan_inf_filterrisnanisinfrxrfloating_point_ops max_grad_normr clip_master_gradsrrgclip_grad_norm_r master_paramsrruget_global_grad_normon_pre_optimizer_steprWon_optimizer_steproptimizer_step_was_skippedReduceLROnPlateau on_step_end_maybe_log_save_evaluateon_substep_endshould_epoch_stopshould_training_stoprr on_epoch_endTPU_METRICS_DEBUG master_printmetmetrics_reportdelattrrbest_model_checkpoint rendezvousrDrfrEdistrro_load_best_modelrd store_flosrr$r'rlog_get_output_dir_sorted_checkpointsrasave_total_limitsamefileshutilrmtreer2_finish_current_pushrSrrX)tXa0b -DND-'  / /9 / M!!22<< ~M Q[\^`oQp   ,1+< (&*&<&< #   y1  & & JJ 4 4SWSuSu 4 v  !3!34 +04::*=$5 "t';';&djjDADJ #&//1##33u<@;K;K;S;SJJ0A0A<8t~t'8YY__!4!4n6L6L M M!--55dnnEDN   .3 3DJ+  "!&D   $ $!//DN " -(()&&(>WefjfpfpWqSq)*d.B.B**+A4CUCUV **+AB 01  23 ' Q'789 o&6q%9:; >tyy?d?def>ghi 99 0 0D4J4J J KK[\`\r\rst[uv w [\rst[uvw 6t7W7W6XYZ 3Ia=AB 9:OPUfj:klm9nop YY[ )*&%)" " -"''.. GGLL/1C D3 &44RWW\\BXZl5mnDJ  4 4TYY K  % % ' !7!7;U!UVN((151G1GKe1f..$2R2RR.12. KK_ ` KK;N;KL M KKA$**BXBXAYZ [(( ,^,<=677SU ; FD D))4t1D E F1A. ++D)=MuU,,s4;;7"%'+zz'='=$ %)  ,,;;D$**dll[    NN5"6tN L>+;<I E/ '5 **51!#! "-$%^^d&F&FF   00??djjRVR^R^_DL&+A+MRptuRu$$%;<KM-1#56FHf#g > 12." D!"23N$t'G'GGIA~ << K*d.N.NNQRRM//14" =)L q BMR_bcRcBdd>>js 484J4J>[fhlhshs4t1 1!*=!9CIAvAID$(1H0P0P#PTU#U#uZ^abZbguYuL$$33GG Uyy>>*1$**>OQ\*]*&8"NN!b ,2/+B+H+H+JL+0<< TYYM]M]ejepep+qL JJ<<@P@P@W@WXd@e@i@i@k@p@p@rr<",,-CD&+ 696!;65A6==a@9Q> 001GH 3?288:592d>>>!C'+'<'<'J'J4QUQ[Q[]a]i]i'j M 2Q 66 ,,==AZAZZ"))$*:*:*B*B%P(33  !]'+'9'9%I['\ ]33 6 8"[[6%++l:S#*Gq4::;Q;Q7QTXTpTp7p,q"q">>\-@-@@","RSZSaSaRbbxzFzMzMyN!O##*L"8%%t/F/Fv/N)OO%#((77KKDQ --9d>P>PST>T68TYY-1^^-M-MdN`N`-a !%-/XX-E-E$'$5$5dnn$E$($6$6." .2-=-=-M-M$)$4$4$6$($6$6." !8 9$($4$4$E$EIbIb$b,1,F,F,H #*9f#=090@I,6 '+'<'<'R'RSWY]YcYceieqeq'r ++-'+'<'<'N'NtUYU_U_aeamam'n )-(?(?(A #//JJ#-d.?.?AYAYAkAk#l $ 1 1 6 6 8) ..!3.+0D1H}4LP^3^+^ ('+'<'<'H'Htzz[_[g[g'h 55#%!!!0&*76 (,'<'<'K'KDRVR\R\^b^j^j'k ||559Z9Z13LLNGCJ<<11T\\5V5V-/ YL Zax ../0##,+-[] 59 100==dDJJPTP\P\]DL  ) )E5%9Mziv * ,, ?)+OOC$6$6$89NN_||00SI V ??wtW5 D' " op  & &4::+K+K+W%' 67##|'?'?? (*  ! ! # 7<<>1 #DJJ$:$:E B,,/DD   )jj**'    $ 5 5  *   44W= &&u-!55RY5Z 99 TZZ%E%E%QVZV_V_VpVptuVu0 B ww'' DJJ4T4TUKK"=j\Ig hiMM*DA B ,,99$ DLLY  !!#  # # /  $ $TZZ 04::11:wGGe  D]]sA_2A_2A_7 _7A` c|j0|-|jtjk(r |j}n|jtjk(r-ddl}|j jj}na|jtjk(r |j}n7|jtjk(rddl }|jj}|j|j|nd}tj j#|j$j&|}|S|j$j&}|S)Nrzrun-)ryrRrnumberrr4r8 get_context get_trial_idridrrrunr"rbrArrr)rr run_idr;rrun_nameros rrDzTrainer._get_output_dir s  ! ! -%2C%%)?)??''?+>+>> ..0==?''?+A+AA''?+@+@@.2ll.Ft||E*dSYRZOHggll499#7#7BGii**Grc  | |j}tjjt}tjjt }tjjt }tjjt}tjjt}tjjt}tjjt} tjjxrmtfdtjDxsDtjjtjjtd} tjjrtjD cgc]} tjjtjj| rtjjtjj| t sBtjjtjj| t r| c} ng} | r|j st#ddtd|||| ||fDs| s| st#dt$j'ddtjj|rLt)j*|} | j,}|)|t.k7r t$j1d |d t.d tjj|s"tjj|s| rt3rtjjtjjd rt5j6td d yt9|j:dr-|j:j<durt$j1dt?j@|dd}d |d<|jC|d}~y|j r@tE|jFjHjJ|jF|fitMy|j:jNrAtjj|r"tPj>jS|d}nt?j@|dd}|jC|d }~|jU|ytW|r+t9|ds t9|drt9|drtjjYrt9|dr5|jZ}t]|dkDrt$j1d|d}n |j^}| rP| D]9}tjj|}|ja||||k(;|jc|y|ja|dyt$j1dt dyt$j1d yte|t3|j:jN!}t3s|jU|yycc} w)"Nc3K|]I}tjjtjj|r t|vKywr)rbrAisdirrFSDP_MODEL_NAME)r folder_namers rrz0Trainer._load_from_checkpoint.. s>77==.Dk!RS ;.sAAz.binzCheckpoint found at z* is only supported when using PyTorch FSDPc3ZK|]#}tjj|%ywr)rbrAr )rfs rrz0Trainer._load_from_checkpoint.. s% q! s)+z!Can't find a valid checkpoint at zLoading model from rSz9You are resuming training from a checkpoint trained with z- of Transformers but your current version is zJ. This is not recommended and could yield to errors or unwanted behaviors.user_content.ptFrAtagrload_optimizerr TzOEnabling FP16 and loading from smp < 1.10 checkpoint together is not supported.r map_location weights_only_smp_is_partialrRractive_adapteractive_adapters load_adapterr zFMultiple active adapters detected will only consider the first adapterr)r>jThe intermediate checkpoints of PEFT may not be saved correctly, consider using a custom callback to save i in corresponding saving folders. Check some examples here: https://github.com/huggingface/peft/issues/96GCould not load adapter model, make sure to have `peft>=0.3.0` installed)rQ prefer_safe)3rrbrArrkrjrirornrmrlr{rRlistdirr r|rJrrrrfrom_json_filetransformers_versionrr<rrorr;rr rloadload_state_dictrrOrprrsave_safetensors safetensors load_file_issue_warnings_after_loadrexistsrr7rr set_adapterr()rrr config_fileadapter_weights_fileadapter_safe_weights_file weights_fileweights_index_filesafe_weights_filesafe_weights_index_file is_fsdp_ckptr}adapter_subdirsrcheckpoint_versionrG load_resultrr subdir_namepeft_ids ` rrzTrainer._load_from_checkpoint s =JJEggll#9;G !ww||,BDXY$&GGLL1GIb$c!ww||$:LI WW\\*@BTUGGLL)?ARS"$'',,/EG^"_ww}}%;<  #%::.D#E  ^ ww~~bggll+AoEVVZC[\] *ww}}34$&::.D#E 77==.Dk!RSGGNN277<<0F Ui#jkww~~bggll3I;Xq&rs     4 434J3KKuvw w  !%&+(-   @AW@XYZ Z )*@)ACD 77>>+ &%44[AF!'!>, '277>>:K+LP\&(77>>"'',,/EGX"YZ..3uej tyy&1diinn6Lm"'Lu[_!`J49J01"'"7"7 4"7"PK"%%$$**66$$*  ,- 99--"''..AR2S!,!2!2!>"89u&78*/*?*?/!3"NN+st)8);).)=)=&+:sK&(ggll3I;&WG!..w S^bpSp.rs)).9**+A>`d*eNNDDXCYZbb hi2-6M6O]a]f]f]w]wK+,// <-} sC [$c tjd|jjd|jjdt j j|jjt}t j j|jjt}t j j|jjt}t j j|jjt}tr |jn |j}|jrAt!|j|jjt#|j y|j$rTt'|j(jj*|j(||jjfit-}yt j j/|s^t j j/|s?t j j/|s t j j/|rd}tr t j j1t j j|jjdr2t3j4|jjtddy|j6j8rAt j j1|r"t:j<j?|d }nt=j@|d d }d|d <|jC|d }yt#|r&tE|ds tE|drtE|drtE|dr=|jFd} tI|jFdkDr"tjKdn |jL} t j j/|st j j/|r7 |jO|jj| ddl-m.} | gg}ntjKdtdd}ntjKdd}n|j6j8rAt j j1|r"t:j<j?|d }nt=j@|d d }|jC|d}ts|r|j_yyyt j j/t j j|jjt`sUt j j/t j j|jjtbrGte||jjt }ts|j_|yytjKd|dy#tP$rR} |jR| jTr3d|jR| jVjXd} tQ| | d} ~ wwxYw)NzLoading best model from z (score: rTrTrFrrrrrrRrrrrr zCDetected multiple active adapters, will only consider the first onez0When using prompt learning PEFT methods such as z, setting load_best_model_at_end=True can lead to errors, it is recommended to set this to False and to load the model manually from the checkpoint directory using PeftModel.from_pretrained(base_model, ) after training has finished.)_IncompatibleKeysrrrz#Could not locate the best model at zi, if you are running a distributed training on multiple nodes, you should activate `--save_on_each_node`.)3rrrpr> best_metricrbrArrormrjrirrNrrCr#rrJrrOrrrr rorrrrrrrrr;rr7r<rrr- peft_configis_prompt_learning peft_typertorch.nn.modules.modulerrrlrnr() rbest_model_pathbest_safe_model_pathbest_adapter_model_pathbest_safe_adapter_model_pathrrhas_been_loadedrGrexcmsgrs rrAzTrainer._load_best_model] s .tzz/O/O.PPYZ^ZdZdZpZpYqqstu'',,tzz'G'GV!ww||DJJ,L,LN_`"$'',,tzz/O/OQe"f')ww||DJJ4T4TVo'p$&=&?""TZZ  $ $ %"" 00'5djj'A#A   ! !)  &&22   00  () K GGNN? +ww~~23ww~~56ww~~:;"O&(77>>"'',,tzz/O/OQb"cd..!ZZ==( %', yy11bggnnEY6Z%0%6%6%@%@AU^c%@%d %*ZZebf%g 49J01"'"7"7 4"7"PK!%( '78GEK\<]cj~d#5*;<-2-B-B1-EN"5#8#89A= &/t u-2-A-AN77>>*ABbggnnUqFr* % 2 24::3S3SUc d R*;B*CK"NN!LL`Kabj!j /4O'pq*/yy11bggnnEY6Z%0%6%6%@%@AU^c%@%d %*ZZebf%g #("7"7 E"JK.0_33K@6E0 WW^^BGGLL)I)IKbc dhjhohohvhv GGLL99;M Ni 2tzz77@W@YK+,// <- NN5o5FGPP e$0 *#(#4#4^#D#W#W)[+0+<+<^+L+V+V+\+\*]^8)8%(+7s*;$D$) *s &Y'' [0A Z==[ct|jdk7r|jjPt |jt |jjk(r|jj n#t jd|jdt|jdk7r$t jd|jdyy)Nrz8There were missing keys in the checkpoint model loaded: rSz;There were unexpected keys in the checkpoint model loaded: ) r7 missing_keysr_keys_to_ignore_on_saver5rrr<unexpected_keys)rrs rrz"Trainer._issue_warnings_after_load s {'' (A -zz11=#kF^F^B_cf 22dC &&(!YZeZrZrYsstuv {** +q 0 NNMkNiNiMjjkl  1rc |j|}|j||jj|t |j t jj jrO|sM|jj}|jdsd|} |j j|||S|S#t$r2}td|dt|jd|d|d}~wwxYw)N ignore_keyseval_9The `metric_for_best_model` training argument is set to 'W', which is not found in the evaluation metrics. The available evaluation metrics are: zX. Please ensure that the `compute_metrics` function returns a dictionary that includes 'zM' or consider changing the `metric_for_best_model` via the TrainingArguments.)evaluater=rprBrrUrr&r2rr startswithrWrrr)rr rrr&metric_to_checkrs rrzTrainer._evaluate s--,@-A !!%)?)?I d'')A)A)S)S T]k"ii==O"--g6$)/):"; !!&&w'?@w OP_O`a==A',,.=Q -C99C>c ||jjrG|jj|jkDr#t rt ji} |j|jj} ||z}t| |jj|jz z d| d<|/t|tjr|jn|| d<||| d<n|j| d<|xj | z c_|jj|_|j#|j%| |d} |jj&r]|j)||} |j+| |} |j,j.t0j2k(r| |j_|jj4rS|j7|||j8j;|j,|j|j|_yy)Nrrr[r)r&r )rw should_logrprBrrrr_nested_gathermeanrrrrrrrrBrCshould_evaluater_determine_best_metricrrrVrra_save_checkpointrYon_save) rrZr[rr r rrrlogstr_loss_scalarr&is_new_best_metrics rr4z Trainer._maybe_log_save_evaluate s << " "tzz'='=@\@\'\%' %'D"009>>@EEGN w G 4::3I3IDLhLh3h!iklmDL$8B9ell8[INN$4aj[!((5_%(,(?(?(A_%  # #~ 5 #+/::+A+AD ( OO  HHT: & << ' 'nnU,@AG!%!>(+ :=/Jccww||J@H77>>(+ B ^ K#(::hT#J  K,X67 09: ""#7#>? ! #   1%8 900L4L4LL :: " " $ $VUZZ9M~ ^ ! # $UEII7K^ \ ! # $UEII7K^ \ ! # $UEII7K^ \ " $ $VUZZ9M~ ^ %# K Ks /JJc 6d}|jjD|jj}|jdsd|} ||}|jj rtjntj}|jj;|jj r tdn td|j_ |||jjrn||j_ |jjtjtj fvr%|jj"|j_d }|S#t$r/}td|dt |j d|d}~wwxYw) z Determine if the model should be saved based on the evaluation metrics. Returns: bool: True if a new best metric was found, else False FNrrrzJ. Consider changing the `metric_for_best_model` via the TrainingArguments.z-infinfT)rrrrrrgreater_is_betterrgreaterlessrprrrrVSTEPSEPOCHrBbest_global_step)rr&r rr metric_valueroperators rrzTrainer._determine_best_metricM s[# 99 * * 6"ii==O"--g6$)/):"; &7 &*YY%@%@rzzbggHzz%%-:>)):U:Uv[`af[g & djj&<&<=)5 &99**|/A/Ar)|jj<| jA| y| |jj<| <|jjCt jj|tD|jjFr|jI||jj,r|jKd|yycc}w)NrR)r Tr?Fr)&rNrprBryrBrDrbrArrCrrrVrrrrr>save_only_model_save_optimizer_and_scheduler _save_scaler_save_rng_staterarYrrwrr4r1r2rrrarDrEr__push_from_checkpoint_rotate_checkpoints) rrr checkpoint_folderrorbest_checkpoint_folderbest_checkpoint_dirrcb_namecb_states rrzTrainer._save_checkpoints sT 55Qtzz7M7M6NO  ! ! )em OO &&U&3WW\\'+<=   48 99 " "|'9'9<;M;M&N NSWS]S]SnSn(='>a @[@[?\%] ""$'',,w8N"O ww~~123F 0yy((  . .z :   j )   , 99 "22<< ~MQ[\^`oQp F,,//88:djj;;GDdKJJ11':AA(K=EDJJ11': F JJ # #BGGLL=O$P Q 99  & &z 2 99   $ $u $ I !s LLc tjtjjtjj d}tj jr~|jjtjk(r,tj jj|d<n+tj jj |d<trtj |d<tr~|jjtjk(r,tj jj|d<n+tj jj |d<t#r~|jjtjk(r,tj$jj|d<n+tj$jj |d<t'r~|jjtjk(r,tj(jj|d<n+tj(jj |d<t+rj|jjtjk(r"tj,j|d<n!tj,j |d<t/j0|d |jj2d kr5t j4|t.j6j9|d yt j4|t.j6j9|d |jj:d y)N)rrrrrrrrrTrr rrr)rgetstater get_stater get_rng_staterrrrDrfrEget_rng_state_allrrrrrrrrrrrbrcrrFrArr)rr rng_statess rrzTrainer._save_rng_state swoo'YY((*<<--/ :: " " $yy&&,*B*BB%*ZZ%6%6%H%H%J 6"%*ZZ%6%6%D%D%F 6" ! # " 0 0 2Ju  ! #yy&&,*B*BB$)II$4$4$F$F$H 5!$)II$4$4$B$B$D 5! ! #yy&&,*B*BB$)II$4$4$F$F$H 5!$)II$4$4$B$B$D 5! ! #yy&&,*B*BB$)II$4$4$F$F$H 5!$)II$4$4$B$B$D 5! " $yy&&,*B*BB%*ZZ%A%A%C 6"%*ZZ%=%=%? 6" J. 99  1 $ JJz277<< O#L M JJz277<< jI`I`Haae  ! !   &&22D4D4DdjjR\ `u`w    &&22D4D4DdnnVZV`V`bl YY " " JJt~~002BGGLL^4\ ])-(A(A) *   8K G % II ! !..2O*,((5 e 4,,779277<< Tb;cd e  0 -3P "Q 5 5X e es4AS6AT6S?T c  y|jrt|jts|t j d5}|jj tjtjjtddddtytr:tjtjjt dzntjj#tjjt xstjj#tjjt$xsHtjj'xr't)fdtj*D}|j,rQtjtjjd|j.j0dt n|}|rtjj#tjjtrt3r|j,rqtjtjjd |j.j4d |j.j0dt d d }|d }n:tjtjjt d d }t j d5}tjtjjtd d }dddtt7j8||j.j:t7j8|j.j:|j<j ||jj |ytrdtjj#tjjdrfd}nfd}|j>jA|n|j.j0dkDr|j.j:nd }|jBrUtE|jFjHjJ|jF|j<|jLfitOnS|j<j tjtjjt |d t j d5}|jj tjtjjtddddtyyy#1swYOxYw#1swYxYw#1swY1xYw)z3If optimizer and scheduler states exist, load them.NTrr_*c3K|][}tjjtjj|rtj dd|v]yw)rSrN)rbrAr{rOPTIMIZER_NAME_BINr)rr}r+s rrz8Trainer._load_optimizer_and_scheduler..1 sM'77==j+)NO+005a8KGsA!A$z rank*-of-rRrrrrrTrc|jtjtjj t dy)NTr)rrorrbrArrHmodoptr+s r opt_load_hookz.opt_load_hook_ s-//jR`9ako0pqrc6trJ|jtjtj j tddy|jtjtj j tdy)NT)r back_compatr)rnrrorrbrArrHrs rr!z.opt_load_hookd sa8 # 3 3$'HHRWW\\*n-U_cqu$v!"!$ 3 3CHHRWW\\*Vd=eos4t urr )(rCrrUrr.rrrrrbrArrIrKrglobrHr rr{rRrrrrrrrsend_cpu_data_to_devicer8rTrNregister_post_step_hookrJrrOrprrr)rr+rcheckpoint_file_existsoptimizer_statelr_scheduler_stater!rs ` rrz%Trainer._load_optimizer_and_scheduler sR     $ $d//1JK,,D9_%%55 277<< N#KZ^_$O4 '( IIbggll:~>E Frww||JGH 77>>"'',,z;M"NO GGMM*-+-::j+A $** IIbggll:499;O;O:PPQR`Qa/bc d'  "bggnnRWW\\*n5]&^%'..&+jj &$tyy/F/F.GtDIIL`L`Kaabcqbr(s&+%) 'O'6k&BO&+jj Z@ucg'O,,D9_). Z@ucg*&$O4**?DII>}M 8> ,, NN JJ&  45 66!JJ " Z HWcrv ,,D9_%%55 277<< N#KZ^_$O4O'_ !9\bs&AW!6;W.8AW;!W+.W8;Xc |jj}|ytrt j dt jd5}t j|jjjtjj|tt|ddd|jj rtst jd5}t#j|jjjtjj|tdddtyyy#t$rYywxYw#1swYxYw#1swY2xYw)Nsaving_scaler_stateTr)rOscalerrrrr?r.rrFrGrbrAr SCALER_NAMErKrrar)rrr,rs rrzTrainer._save_scaler s+ %%,,F >  ! # MM/ 0((5 5((//::cd#O4 5 99 )?)A((5 h 4++22==?jZeAfg h  0*B      5 5  h hs+E%A&E44AF% E10E14E=F c L|ytjjtjj|t}|rEt rt jd5}tjtjj|tdd}dddttj|jj|jj j#|yt jd5}|jj j#tjtjj|tddddtyy#1swYxYw#1swY"xYw)z If scaler state exists, load it.NTrrrr)rbrAr rr-rr.rrrrKrr%rr8rOr,r)rr+r'r scaler_states rr zTrainer._load_scaler s,   !# Z0U!V !&',,D9_#(:: Z=E`d$L$O4**<9I9IJ  ''77 E,,D9_$$++;; 277<< K#HW[\$O4! "s(;FAFFF#c |jjsyg}g}|jj|jgz}|j j jD]9\ }t|ts|g}t fd|Dr|Dcgc]}|jj k(s| }}t||D]\}}|jdi}|jdi} t|d i|} | jD]\} } t!| | | t|t"r| |_n|j%| |jj't| t(j+d)|j% <t-|dkDr(t(j/ddj1|d |D]}|jj3|ycc}w) zLIf callback states exist and were passed in, restore their states if enabledNc3PK|]}|jjk(ywr)r1r2)rrstored_callbacks rrz/Trainer._load_callback_state.. s"eh8%%../Aes#&r attributeszPContinuing training from checkpoint, restoring any callbacks that were passed inrzPCheckpoint included callbacks not included in current configuration. Ignoring. (rrr)r'restore_callback_states_from_checkpointrYrrwrprrrrrRr1r2ziprrrr8rarrrr7r<rrZ)r not_found new_callbacksoriginal_callbacksrhr duplicates callback_datarr3 new_callback attributerr2s @rr zTrainer._load_callback_state syy@@   !22<< ~M%)ZZ%B%B%H%H%J 2 !OTdD)veRdee.@!)8CUCUC^C^bqCqH 03:t/D N+Hm(,,VR8D!.!2!2<!DJ#14>#9D#9L,6,<,<,>@( 5 i?@!(N;'3 %,,\:))99$|:LM N no  1/ 20 y>A  NNbcgclclmvcwbxxyz & 9H  ! ! . .x 8 9/s G;6G;rr-n_trials directionbackendrr"c 8| t}t|}t|}|j||_|j t d| |jn||_||_ |tn||_ |j|||fi|} d|_| S)az Launch an hyperparameter search using `optuna` or `Ray Tune` or `SigOpt`. The optimized quantity is determined by `compute_objective`, which defaults to a function returning the evaluation loss when no metric is provided, the sum of all metrics otherwise. To use this method, you need to have provided a `model_init` when initializing your [`Trainer`]: we need to reinitialize the model at each new run. This is incompatible with the `optimizers` argument, so you need to subclass [`Trainer`] and override the method [`~Trainer.create_optimizer_and_scheduler`] for custom optimizer/scheduler. Args: hp_space (`Callable[["optuna.Trial"], Dict[str, float]]`, *optional*): A function that defines the hyperparameter search space. Will default to [`~trainer_utils.default_hp_space_optuna`] or [`~trainer_utils.default_hp_space_ray`] or [`~trainer_utils.default_hp_space_sigopt`] depending on your backend. compute_objective (`Callable[[Dict[str, float]], float]`, *optional*): A function computing the objective to minimize or maximize from the metrics returned by the `evaluate` method. Will default to [`~trainer_utils.default_compute_objective`]. n_trials (`int`, *optional*, defaults to 100): The number of trial runs to test. direction (`str` or `List[str]`, *optional*, defaults to `"minimize"`): If it's single objective optimization, direction is `str`, can be `"minimize"` or `"maximize"`, you should pick `"minimize"` when optimizing the validation loss, `"maximize"` when optimizing one or several metrics. If it's multi objectives optimization, direction is `List[str]`, can be List of `"minimize"` and `"maximize"`, you should pick `"minimize"` when optimizing the validation loss, `"maximize"` when optimizing one or several metrics. backend (`str` or [`~training_utils.HPSearchBackend`], *optional*): The backend to use for hyperparameter search. Will default to optuna or Ray Tune or SigOpt, depending on which one is installed. If all are installed, will default to optuna. hp_name (`Callable[["optuna.Trial"], str]]`, *optional*): A function that defines the trial/run name. Will default to None. kwargs (`Dict[str, Any]`, *optional*): Additional keyword arguments for each backend: - `optuna`: parameters from [optuna.study.create_study](https://optuna.readthedocs.io/en/stable/reference/generated/optuna.study.create_study.html) and also the parameters `timeout`, `n_jobs` and `gc_after_trial` from [optuna.study.Study.optimize](https://optuna.readthedocs.io/en/stable/reference/generated/optuna.study.Study.html#optuna.study.Study.optimize) - `ray`: parameters from [tune.run](https://docs.ray.io/en/latest/tune/api_docs/execution.html#tune-run). If `resources_per_trial` is not set in the `kwargs`, it defaults to 1 CPU core and 1 GPU (if available). If `progress_reporter` is not set in the `kwargs`, [ray.tune.CLIReporter](https://docs.ray.io/en/latest/tune/api/doc/ray.tune.CLIReporter.html) is used. - `sigopt`: the parameter `proxies` from [sigopt.Connection.set_proxies](https://docs.sigopt.com/support/faq#how-do-i-use-sigopt-with-a-proxy). Returns: [`trainer_utils.BestRun` or `List[trainer_utils.BestRun]`]: All the information about the best run or best runs for multi-objective optimization. Experiment summary can be found in `run_summary` attribute for Ray backend. NzXTo use hyperparameter search, you need to pass your model through a model_init function.) r rRrensure_availableryrr-default_hp_spacerr"rZr-rv) rrr-r=r>r?r"r backend_objbest_runs rhyperparameter_searchzTrainer.hyperparameter_search s@ ?/1G!'*8AC $$&!( ?? "j 9A8H 44h  >O>W!:]n";??49GG!%rrrc |jj|jj|d<|jjr=|jj|d<|"t d||jji|d|jj i}|jjj||jj|j|j|j||_ y)a8 Log `logs` on the various objects watching training. Subclass and override this method to inject custom behavior. Args: logs (`Dict[str, float]`): The values to log. start_time (`Optional[float]`): The start of training. Nr rr8)r rW) rpr rrrrdrB log_historyrarYon_logrw)rrrrs rrCz Trainer.log/s ::   ' JJ,,DM 99 2 2,0JJ,L,LD( )%gzdjj>^>^_=D=VTZZ%;%;<= %%f-,,33DIItzz4<.MsCd11!4Cr8rr)rrrrrKrrrrrr8rCis_floating_point is_complexrbrOrpr!rrr)rrhrr#rs` rrKzTrainer._prepare_inputFs dG $4:TZZ\RTQq$"5"5a"88RS S udm ,4:CdCC C ell + 0 01F((e.E.Ed.KuO_O_`dOe w(8(8(>(>(O(O(\(\(b(b(def477$V$ $ SsE rhc|j|}t|dk(r(tddj|jd|j j dk\r|j|j|d<|S)z Prepare `inputs` before feeding them to the model, converting them to tensors if they are not already and handling potential state. rzThe batch received was empty, your model won't be able to train on it. Double-check that your training dataset contains keys expected by the model: rGrSmems)rKr7rrrjrrrrrhs rrVzTrainer._prepare_inputsXs $$V, v;! IILRVRiRiIjHkkln  99  1 $)?!ZZF6N rc"|jS)zF A helper wrapper to group together context managers. )autocast_smart_context_managerrs rcompute_loss_context_managerz$Trainer.compute_loss_context_managerhs2244rrNc|jr7tjjj ||j }|St j}|S)z A helper wrapper that creates an appropriate context manager for `autocast` while feeding it the desired arguments, depending on the situation. )rNr)rlrrrrXrtrr)rrN ctx_managers rrSz&Trainer.autocast_smart_context_managernsO   ))--00}TXTbTb0cK%002Krcz|jt|jdr9t|jjr|jj|j |}t rbt |||jj}|jjj|jjS|j5|j|||}ddd~|jj>|j j"|jjzdk(r t%rt&j(j+nt-rt&j.j+nt1rt&j2j+nt5rt&j6j+nit9drt&j:j+n>t=rt>jAdnt&jBj+i}|jjDtFjHtFjJfvr|jM|d<|jjNd kDrjQ}|jRr:tUjV|j5}|jYdddy|jZs%|j\|jjz }|j^j`tbjdk(rd |d <|j^jXfi||jS#1swYxYw#1swYyxYw) aq Perform a training step on a batch of inputs. Subclass and override to inject custom behavior. Args: model (`nn.Module`): The model to train. inputs (`Dict[str, Union[torch.Tensor, Any]]`): The inputs and targets of the model. The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument `labels`. Check your model's documentation for all accepted arguments. Return: `torch.Tensor`: The tensor with training loss on this batch. r8rfNrz2.0) min_versionzW`torch_empty_cache_steps` is set but HPU device/backend does not support empty_cache().rr F scale_wrt_gas)3r8r;rTrdrVrrrr reduce_meandetachrr8rT compute_losstorch_empty_cache_stepsrprBrrxpu empty_cacherrrrrrrmpsrrr<rr&rerrrrurrkr scale_lossbackwardrQrrOr"rr#)rrrhrfloss_mbrr scaled_losss rr%zTrainer.training_stepzs(  4>>7 +9M9M0N NN "%%f- " $*5&$)):_:_`G&&(//144TYY5E5EF F  . . 0 [$$UFGY$ZD [  II - - 9 &&)J)JJaO%' %%'') %%'(* &&(') %%''E: %%'')m &&( 99??~22N4J4JK K&*&=&=&?F? # 99??Q 99;D ==dnn5 '$$& ' '11d6L6L6TdiiCCC00O4M4MM*/' %D   % %d 5f 5;;= c [ [F ' 's-N$N1$N.1N:c $|j |jd|vr|jd}nd}|jri}|||d<i||}|d i|}|jj dk\r||jj |_||jj|}t|r%|jjj} n|j} |j|j|||} n| tjvr|j||d} n|j||} nzt|t rPd|vrLt#d d j%|j'd d j%|j'd t|t r|dn|d} |jj(r3|js |jr|| |jj*z} |r| |fS| S)z How the loss is computed by Trainer. By default, all models return the loss in the first element. Subclass and override for custom behavior. NlabelsrfrrXT) shift_labelsrzJThe model did not return a loss from the inputs, only the following keys: rGz,. For reference, the inputs it received are rSr)rvrr$rQrrrrOr)rrr _get_namer*r6rrrrraverage_tokens_across_devices num_processes) rrrhreturn_outputsrfrg loss_kwargsoutputsrrrs rr]zTrainer.compute_losss     +t/E/E/QW_ciWiZZ)FF  ) )K!-4F 01..+.F/&/ 99  1 $ !5!56DJ  "..;;EBOo.,77==GGI ,668 %%1--gvRd-e@GGII**7F*N**7F;'4(V7-B `xx /00\]`]e]eflfqfqfs]t\uuvx '1$&?76?WQZD II 3 3//43I3I". D$$22 2D"0g:d:rc4|jjdk(S)z Whether or not this process is the local (e.g., on one machine if training in a distributed fashion on several machines) main process. r)rlocal_process_indexrs rrzTrainer.is_local_process_zeros yy,,11rcvtrtjdk(S|jjdk(S)z Whether or not this process is the global main process (when training in a distributed fashion on several machines, this is only going to be `True` for one process). r)rrorrrrs rrzTrainer.is_world_process_zeros/ # $88:? "99**a/ /rrr@c||jj}tr|j|nt rt j |d|jj}|jjr|j||tr|tt jj|djnC|j rdt#|j$j&j(j*vrt-j.t0t-j.dkDr|j$j3|j4}|jjr|j||n|j6rP |j$j3|j8}|jjr|j||n'|jjr|j||jjHr|s|jId yyy#t:$rt<j?d|jjr|j|itA|jj|tBtDg|jjG|YwxYw) z Will save the model, so you can reload it using `from_pretrained()`. Will only save from the main process. NTr)rGrFULL_STATE_DICTz0.24.1z| stage3_gather_16bit_weights_on_model_save=false. Saving the full checkpoint instead, use zero_to_fp32.py to recover weightsz Model save)commit_message)%rrr _save_tpurrbrcrNrGra_savernrrArtouchrJrrOrprstate_dict_typerraccelerate_versionget_state_dictrrCr#rrr<rLrormrr_)rrr@rGs rrCzTrainer.save_models  --J ! # NN: & $ & KK T 2++668Jyy$$ :* =(RWW\\*.?@AGGI  ! !!S)9)9)?)?)K)K)[)[%\\ 01GMM(4KK!--< ?s'AI B KKc ||n|jj}tjd||j}t j t jdrYtj|dtj|jtjj|ttf}t j d|j"r|j%|j'd}tjj|d|jj(d |jj*d t,}t j||d t j d |jj.rd dlm}|tjj|ddt,d\}}|j4j4}|j6j9|} t;| |r9| j=||t j|jj>ntjdt j|tjj|t,nt;||st;|j6j9||r|j6j9|j=||jj.t j@|j%t j|jj>ntjdt j@|j%} t j| tjj|t,nm|j=||jj.t j|jj>t j@|j%|jB3|jj.r|jBj=|yyy)NSaving model checkpoint to F)localTrsaving_checkpoint)rrrrrRrsave_full_checkpointsr)%consolidate_sharded_model_checkpointsrz rank*-of-*-) ckpt_prefix ckpt_suffixrC)rG save_functionsafe_serializationETrainer.model is not a `PreTrainedModel`, only saving its state dict.)is_main_processrGrr)rrrrG)"rrrrrrris_master_ordinalrbrcrrFrArTRAINING_ARGS_NAMErrr?rrGr rrrorarrr9rOr)rsave_pretrainedr_maybe_convert_to_cpur) rrrsupported_classesckpt ckpt_pathrfull_state_dictrQrrGs rruzTrainer._save_tpuDs:#-#9Ztyy?S?S  1*>?    e , KK T 2 JJtyy"'',,z;M"N O,- )*  & &))+"'":":"<D d499#:#:";4 @T@T?UUVWcVdeI GGD) 7 MM1 2yy$$\%J " Z <"-l^ <$&"  ++"&"2"2"?"?"Fo/@A#33"#2&(gg+/99+E+E 4KK ghGGORWW\\*l-STE#45$**77>@QR  --e4DD$(II$9$9!778H8H8JK"$'''+yy'A'A E cd55e6F6F6HI  BGGLL\$JK  ! ! $ 5 5 gg#'99#=#=33E4D4D4FG "   ,1F1F  ! ! 1 1* =2G ,rc||n|jj}tj|dtj d|t stfn ttf}t|j|sB||jj}t|jj|j|rL|jj|jj|||jjntj d|jjrGt j"j%|tj&j)|t*ddinkt#j,|tj&j)|t.n2|jj|||jj|j0|j0j|nr|j2ft5|j2d rP|j2j6:tj d |j2j6j|t#j,|jtj&j)|t8y) NTrr|)rGrrrpt)rrzWSaving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`)rrrbrcrrrr'rrrrGrOr)rrrr save_filerArrmrFrorrr;rr)rrrGrs rrvz Trainer._saves #-#9Ztyy?S?S  J. 1*>?6G6I_.P_ajOk$**&78!!ZZ224 $**77 CEVW  --djj9II:$))JdJdJ cd99--%%//"BGGLL=N$O[ceiZj0JJz277<< L+QR JJ & &zdiiF`F` '   ,  ! ! 1 1* =    ***K8"",,8 KKq r    ( ( 8 8 D  499bggll:7IJKrc|jjtjk(rm|jxj t |jg|jjjjz c_d|_y|jxj |jz c_d|_y)Nrr) rrDrfrErprrArxr8r+rrs rrBzTrainer.store_floss 99 " "l&>&> > JJ ! !-t/@/@.A$))JZJZ[__affh  !!"D  JJ ! !T%6%6 6 ! !D rc~g}t|j|dDcgc]-}tjj |s#t |/}}|D]}|r1|j tjj||f6tjd|d|}|S|jd|j t|jd|ft|} | D cgc]} | d } } |jjt t|jj| vrk| jt t|jj} t!| t#| dz D]} | | dz| | c| | <| | dz<| Scc}wcc} w)N-*z.*z -([0-9]+)rr r)rr$rbrAr{rragetmtimerematchgroupsrsortedrpr>indexrr7) rrcheckpoint_prefixrordering_and_checkpoint_pathxglob_checkpointsrA regex_matchrpr+best_model_indexrgs rrEzTrainer._sorted_checkpointss(*$,0,<,A,AEVDWWYBZ,[pq_a_f_f_l_lmn_oCFpp$ ^D,33RWW5E5Ed5KT4RS hh,=+>i'H$O *{/A/A/C/O077[=O=O=QRS=T9UW[8\]  ^$$@A>PQ jmQQ JJ , , 8D99:;?QQ177DAaAa|jj|jjdkry|j||}t||jjkry|jj}|jj 7|jjdk(r|d|jj k7rd}t dt||z }|d|}|D]2}tjd|dtj|d 4y) Nrrr rrrrTr) rrFrEr7rpr>rtrrrHrI)rrrrprFnumber_of_checkpoints_to_deletecheckpoints_to_be_deletedr+s rrzTrainer._rotate_checkpointss 99 % % -1K1Kq1P "55 V`5a ! "dii&@&@ @  9955 JJ , , 8 **a/"2&$***J*JJ *-a5G1HK[1[*\'$67W8W$X!3 :J KK5j\A_` a MM*D 9 :rrmetric_key_prefixc z|du}|r|n |j}t|trJi}|jD]3\}}|j |r|n|||d|}|j |5|S|j j|j|} |jr t| } tj} |jjr |jn |j} | | d|j dnd||} |jj"|jj$z} |d| j&vr| | j&|dz } |d| j&vr| | j&|dz } | j&j t)|| | j*t-j.| j*| z  |j1| j&t2j4|jj6vr't9j:t=j>|j@jC|j|jD|jF| j&|_#|j jI| j&| j&S) a Run evaluation and returns metrics. The calling script will be responsible for providing a method to compute metrics, as they are task-dependent (pass it to the init `compute_metrics` argument). You can also subclass and override this method to inject custom behavior. Args: eval_dataset (Union[`Dataset`, Dict[str, `Dataset`]), *optional*): Pass a dataset if you wish to override `self.eval_dataset`. If it is a [`~datasets.Dataset`], columns not accepted by the `model.forward()` method are automatically removed. If it is a dictionary, it will evaluate on each dataset, prepending the dictionary key to the metric name. Datasets must implement the `__len__` method. If you pass a dictionary with names of datasets as keys and datasets as values, evaluate will run separate evaluations on each dataset. This can be useful to monitor how training affects other datasets or simply to get a more fine-grained evaluation. When used with `load_best_model_at_end`, make sure `metric_for_best_model` references exactly one of the datasets. If you, for example, pass in `{"data1": data1, "data2": data2}` for two datasets `data1` and `data2`, you could specify `metric_for_best_model="eval_data1_loss"` for using the loss on `data1` and `metric_for_best_model="eval_data2_loss"` for the loss on `data2`. ignore_keys (`List[str]`, *optional*): A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions. metric_key_prefix (`str`, *optional*, defaults to `"eval"`): An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named "eval_bleu" if the prefix is "eval" (default) Returns: A dictionary containing the evaluation loss and the potential metrics computed from the predictions. The dictionary also contains the epoch number which comes from the training state. NrQ)rrr EvaluationT)rprediction_loss_onlyrr_jit_compilation_time_model_preparation_timerr)%rrrrrrbr'r(rrr%rrrprediction_loopevaluation_looprrrr&rdrmathceilrCrr9r/rr:r;r<rY on_evaluaterprwr)rrrroverrider&eval_dataset_name _eval_datasetdataset_metricsrr eval_looprtotal_batch_sizes rrzTrainer.evaluatesrZ t+'/|T5F5F lD )G4@4F4F4H 00!="&--2:@Q +):(;1=N>S)TU UJ 7 8FNN J &..,=+>>U)VW WJ !"..))F$6$69I$IJ      ( (DIIOO ; OOC..0 1,,88DJJPTP\P\^d^l^lm  44V^^D~~rc |jj|j|}tj}|jj r |j n |j}||d||}|jj|jjz}|d|jvr||j|dz }|d|jvr||j|dz }|jjt|||jtj|j|z |j j#|j|j$|j&|j|_|jj)|jt+|j,|j.|jS)a Run prediction and returns predictions and potential metrics. Depending on the dataset and your use case, your test dataset may contain labels. In that case, this method will also return metrics, like in `evaluate()`. Args: test_dataset (`Dataset`): Dataset to run the predictions on. If it is an `datasets.Dataset`, columns not accepted by the `model.forward()` method are automatically removed. Has to implement the method `__len__` ignore_keys (`List[str]`, *optional*): A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions. metric_key_prefix (`str`, *optional*, defaults to `"test"`): An optional prefix to be used as the metrics key prefix. For example the metrics "bleu" will be named "test_bleu" if the prefix is "test" (default) If your predictions or labels have different sequence length (for instance because you're doing dynamic padding in a token classification task) the predictions will be padded (on the right) to allow for concatenation into one array. The padding index is -100. Returns: *NamedTuple* A namedtuple with the following keys: - predictions (`np.ndarray`): The predictions on `test_dataset`. - label_ids (`np.ndarray`, *optional*): The labels (if the dataset contained some). - metrics (`Dict[str, float]`, *optional*): The potential dictionary of metrics (if the dataset contained labels). Prediction)rrrrrr) predictionsrr&)r'r(r rrrrrrrr&rbrdrrrrY on_predictrprwrrTrr) rrrrtest_dataloaderrrrrs rpredictzTrainer.predict^sH ""$22<@YY[ ,0II,P,PD((VZVjVj  ;bs  9944tyy7K7KK 5 6&.. H &..,=+>>S)TU UJ 7 8FNN J &..,=+>>U)VW WJ !"..))F$6$69I$IJ   ,,77 4::t||]c]k]kl  44V^^DF,>,>&JZJZdjdrdrssrrc L|j}||n |j}|jr|jt |dd\}}|j |j d|}t|jjdk(r||j urtj} |js%|jr4|jjdk7r|jj|n|jj|d}ttj| z d |_|jr||_||j ur||_|jr|j"|_|j$so|j&r,|j)t*j,|j. }n7|j0r+|j)t*j2|j. }|jj4} t6j9d |d t;|r(t6j9d |j=|nt6j9dt6j9d| |j?tA|jBdr9tE|jBj>r|jBj?||jF_$tK|dd} |jLdk\rd|_'tQ|jjRd} tQ|jjRd} tQ|jjRd}tQ|jjRd}d}i}d}tU|D]j\}}tW|}| ||z }| |} |jY||||\}}}tK|j dd}d|jZvr|j]||nd}t_rtajb|1|je|jg| }| ji||[|jjk|dd}|je|}|jjlr|dk(r|ji|||jjk|dd}|y|jjk|dd}|jn|jo||}|je|}|jjlr|dk(r| ji||=|je|}|jjlr|dk(r|ji||jFjq||jr|jt|_:|jjlr|jvo|m|k|jjxjz}i}d|jZvr|nd|d<d|jZvr|nd|d<|jwt}d(||d||}~~~~t*j~j|j|dz|jzdk(s | j| j|j|j~~~~t*j~jm|jj|_2|jLrtA|dr t|d| j} | j} |j}|j}t;| r t| }nLt| trtK| d ddkDr | j<}nt;|r|j=|}n|}|dk(r|dkDr|}|jvb| `|^|jjlsHd|jZvr| nd|d<d|jZvr|nd|d<|jwt}d(| |d|}n|i}t|}t| tr:| r8tj| jj||d!<n>t| tjr$| jj||d!<tA|d"r|j||d#<tA|d$r|j ||d%<t|jD]0}|j|d&r|j|||d&|<2t| |||'S)) Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`. Works both with or without labels. NrTr  inferenceFrrrevaluation_moderrr8 ***** Running  *****rz Num examples: Unknown Batch size = rr padding_indexrrrrhr )dim pad_indexrrlossesrrrrr_lossrrmodel_preparation_timerrQrrr&rr)TrrrCr#r"rrr7rO_modelsrrJrr prepare_modelrrrNr$rGrrfloat16r8rHrsrrrr_rrr;rTrdrYrr3rrr;eval_do_concat_batchesrrCprediction_stepinclude_for_metricsrKrrrgather_functionrepeatrpad_across_processesrron_prediction_steprprwrrend_of_dataloaderrQrr`eval_accumulation_stepsto_cpu_and_numpygather_for_metricsr= get_arraysrr<r[rr concatenaterrrrrrr$rP)rrrrrrrrQrrrr all_losses all_preds all_labels all_inputsr&eval_set_kwargsobserved_num_examplesrWrhobserved_batch_sizerrrgr inputs_decode is_last_step batch_kwargsrrs rrzTrainer.evaluation_loopsyy7K7W3]a]v]v  $ $)?!$1MDAq  e  S t'' (A -%4::2EJ,,1E1E$JZJZJjJjnsJs  ((/%%33E43P  +0 j0H!*LD '##" DJJ&%*"((!%!3!3""u}}T[[I$$u~~dkkJYY..  &{m6:; j ! KK+D,=,=j,I+JK L KK1 2 oj\23 4>>6 *x8K8K/L NN   !0:-z9d; ??a DJ'tyy'G'GW[\ %dii&F&FVZ[ &tyy'G'GW[\ &tyy'G'GW[\ !"&j1C )LD&"1&"9 ".%)<<%%!4J&*%9%9%I]kv%9%w "FFF%djj2C[QO@HDLdLd@d##F?$;>v1X\>]!))>>v1X\>]55A!??OF--f5yy33{l7RMM&)!--f5yy33{l7RNN6*00CCD$**VZVbVbcDLyy++''38JvOa#'#3#3#B#B#T#TL#%L7=AYAY7YV_cL*7?4C[C[7[VaeL*"22&\6V\|\'33G FFF &&(--9tax4KgKg>gkl>l++-**,++-++-FFF &&(GC )L $//BB ??wtW5 D' " **, ((* **, **,  l #l+K &: ; Vdfg@hkl@l&33K*%"// ; 3 !  5 9/K  ,%&II006<@X@X6X ^bOH %6>$BZBZ6Z `dOH %**^9 ^o^G_G(0 j$ 'J35>>*3M3R3R3T3Y3Y3[G()/ 0  BJJ /3=??3D3I3I3KG()/ 0 4/ 0CGC\C\G())>? @ 41 2EIE`E`G())@A B ' IC>>%6$7q"9:8? C8H,-Qse45 I)zSZhsttrcX|ytr|d}t||}|Str t|}|S|jj #|jj j dk7s/|jj $|jjdk7r t|}|S) Gather value of `tensors` (tensor or list/tuple of nested tensors) and convert them to numpy before concatenating them to `gathered` N nested_gatherNOr) rrJrrrdistributed_stater" local_rankrBrtensorsnames rrzTrainer._nested_gatherts ?  ! #|&,Wd;G % & )G  ii))5$)):U:U:f:fjn:n II ' ' /DII4H4HB4N(1Grct|jdk(rdntfd|jD}jdd}| |j}t|jdk(r|rdnd}|j ;t |jdr#t|jjdd gng|s|r;ttfd |jD}t|d k(r|d}nd}tj5trt|} |s|r{t!| t"r(| d } tfd | j%D} n | d} | d d} | j'j)j+} t-| } nad} t!| t"r#tfd| j%D} n| } t-| } n|s|r|j/5|j1|d\} }ddd j)j3} t!t"r#tfd|j%D} n|d d} nd} |j/5|di}dddt!t"r#tfd|j%D} n|} |j4j6dk\r!||j4j6d z |_ddd|r ddfSt } t| d k(r| d}  | |fS#1swY%xYw#1swYxYw#1swYLxYw)a Perform an evaluation step on `model` using `inputs`. Subclass and override to inject custom behavior. Args: model (`nn.Module`): The model to evaluate. inputs (`Dict[str, Union[torch.Tensor, Any]]`): The inputs and targets of the model. The dictionary will be unpacked before being fed to the model. Most models expect the targets under the argument `labels`. Check your model's documentation for all accepted arguments. prediction_loss_only (`bool`): Whether or not to return the loss only. ignore_keys (`List[str]`, *optional*): A list of keys in the output of your model (if it is a dictionary) that should be ignored when gathering predictions. Return: Tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[torch.Tensor]]: A tuple with the loss, logits and labels (each being optional). rFc3DK|]}j|duywrr)rrrhs rrz*Trainer.prediction_step..s Av`a&**Q-W[B[Avs  return_lossNTrkeys_to_ignore_at_inferencepast_key_valuesc3@K|]}j|ywrr)rrrhs rrz*Trainer.prediction_step..s(WdD)9(WrLr rc3:K|]\}}|dgzvs|ywrNrrrr#rs rrz*Trainer.prediction_step..s&)o1qXcgmfnXnOn!)oc32K|]\}}|vs |ywrrrs rrz*Trainer.prediction_step..s)d1qXcOc!)d )rlc3:K|]\}}|dgzvs|ywrrrs rrz*Trainer.prediction_step..s&&hTQQ\`f_gQgHgq&hrc32K|]\}}|vs |ywrrrs rrz*Trainer.prediction_step..s&]TQQ\H\q&]rr)r7rzallrrsrVr;rr3rrHrrrYrrrrrr[r\rrrTr]rrrr)rrrhrr has_labelsrloss_without_labelsrg raw_outputsrd logits_mbrrrns ` ` rrzTrainer.prediction_steps<"$"2"23q8UcAveieueuAv>v jj5  ..K&)$*:*:&;q&@[dV[%%f-  tzz8,%djj&7&79VYjXkl   ,"5(WdFVFV(W#WXF6{aF ]]_( G&(.uf= !4!+t4"-f"5$))o 8I8I8K)o$o "-a.$/O "..0779==?D.y9FD!+t4$))d 8I8I8K)d$d $/ .y9F!4::<^(,(9(9%X\(9(] g^;;=--/D!'40!&&hW]]_&h!h!(D::<2"'/&/2!'40!&&]W]]_&]!]!(yy++q0%,TYY-A-AA-E%F Q( GT $% %v& v;! AYFff%%7^^22A( G( Gs?C.M M"A1M MA7MM MM MM'cft|jdr|jj|Sy)a For models that inherit from [`PreTrainedModel`], uses that method to compute the number of floating point operations for every backward + forward pass. If using another model, either implement such a method in the model or subclass and override this method. Args: inputs (`Dict[str, Union[torch.Tensor, Any]]`): The inputs and targets of the model. Returns: `int`: The number of floating-point operations. r)r)r;rr)rQs rr)zTrainer.floating_point_opss+ 4::3 4::008 8rtokenc|jsy|jj8t|jjj j }n|jj}||n|jj}t|||jjd}|j|_d|_ y)zE Initializes a git repo in `self.args.hub_model_id`. NT)r privater) rrr^rrabsoluter hub_tokenr hub_private_reporepo_idpush_in_progress)rr  repo_namerepo_urls rr`zTrainer.init_hf_repos ))+  99 ! ! )TYY112;;=BBI ..I* 0C0Cytyy?Y?Ydhi$,, $rlanguagelicensetagsrfinetuned_fromtasks dataset_tags dataset_argsc H|jsytjj|jj d} d} tjj | rtj| jjd} | dk(} tj| jj} |2| 0t|tr|g}| D]}||vs|j|tj ||||||||||  }|j#}t%| d5}|j'|ddd| rI|j(j+|j,j/|jj yy#1swYUxYw)a Creates a draft of a model card using the information available to the `Trainer`. Args: language (`str`, *optional*): The language of the model (if applicable) license (`str`, *optional*): The license of the model. Will default to the license of the pretrained model used, if the original model given to the `Trainer` comes from a repo on the Hub. tags (`str` or `List[str]`, *optional*): Some tags to be included in the metadata of the model card. model_name (`str`, *optional*): The name of the model. finetuned_from (`str`, *optional*): The name of the model used to fine-tune this one (if applicable). Will default to the name of the repo of the original model given to the `Trainer` (if it comes from the Hub). tasks (`str` or `List[str]`, *optional*): One or several task identifiers, to be included in the metadata of the model card. dataset_tags (`str` or `List[str]`, *optional*): One or several dataset tags, to be included in the metadata of the model card. dataset (`str` or `List[str]`, *optional*): One or several dataset identifiers, to be included in the metadata of the model card. dataset_args (`str` or `List[str]`, *optional*): One or several dataset arguments, to be included in the metadata of the model card. Nz README.mdF library_namer) rrrrrrrrrw)rrbrArrrrr rrhrrrrrar& from_trainer to_model_cardopenwriterOr)rcreate_or_update_model_card)rrrrrrrrrrmodel_card_filepathis_peft_libraryr existing_tagsrtraining_summary model_cardrs rcreate_model_cardzTrainer.create_model_cardswJ))+  ggll499+?+?M 77>>- .$>>*=>CCGGWL*f4O&NN+>?DDIIMM$=dC( 6D()C$ C()+77 !)%%  &335 %s + q GGJ       ) )$** 5 Q QRVR[R[RfRf g   s 2FF!c |jr'|jjtjk(ry|jj s'|j |j jsy|jj}tttg}ttfD]}tjj!||}tjj#|sC|j%|t'|5}t)j*|j-}dddt/t1dj3}|j5|t7r |j5t8t:t<g|D]} tjj#tjj!|| sAt?j@tjj!|| tjj!|| |jB|jBjE|tGjH|jtjj!|tJ|jjLtNjPk(rd|jRjT} n"dtW|jRjX} t[|j\|| |jj^ddt`dg} | g} |jjtjbtjdfvr|jjtjbk(rdntg|jh} t[|j\|| | d z|jj^d }| j%||j |j jrtk| |_y|j jlj5| y#1swYxYw) N weight_mapzTraining in progress, step zTraining in progress, epoch Trr)r folder_pathrtr  run_as_futureignore_patternszlast-checkpointz , checkpoint)rr+ path_in_reportr r,)7rr hub_strategyrSENDhub_always_pushris_donerrkrormrnrlrbrArr rar jsonloadsreadrr5r6extendrrhrjrirHr,rrrrFrrrVrrprBrr rr^rrN CHECKPOINTALL_CHECKPOINTSrrrqjobs)rrrmodeling_files index_file index_pathrr shard_files modeling_filertmodel_push_job push_jobsr.checkpoint_pushs rrzTrainer._push_from_checkpoint^sB))+tyy/E/E/X yy((T-B-B-NW[WlWlWtWtWv YY)) %|5FG-/FG 3J&7DJww~~j)%%j1*%1 JJqvvx0E1"3u\':'A'A'C#DE %%k2 3    ! !#68LNg"h i+ uMww~~bggll+Y>Y%Z Z%)YY%;%;{?U?U%U![_`q[r[w[w ,))-)->ii))" O   _ -  (D,A,A,I,I,K$29$=D !  ! ! & & - -i 8a11s $QQ ct|dsy|jK|jjs0tj d|jj yyy)Nrz\Waiting for the current checkpoint push to be finished, this might take a couple of minutes.)r;rr2rrwait_until_doners rrJzTrainer._finish_current_pushsUt/0   ,T5J5J5R5R5T KKv w  ! ! 1 1 36U ,rrtblockingrevisionc ~|jdd}|~|jjrh|jj*t |jj j }n(|jjjdd}||n|jj}|j|j||jd|jsyt|jdd\d |vrg|d <t|d tr |d g|d <|jj D]}||d vs |d j#| |j$d d|i||j't)|j|jj ||| d t*d g| S)u Upload `self.model` and `self.processing_class` to the 🤗 model hub on the repo `self.args.hub_model_id`. Parameters: commit_message (`str`, *optional*, defaults to `"End of training"`): Message to commit while pushing. blocking (`bool`, *optional*, defaults to `True`): Whether the function should return only when the `git push` has finished. token (`str`, *optional*, defaults to `None`): Token with write permission to overwrite Trainer's original args. revision (`str`, *optional*): The git revision to commit from. Defaults to the head of the "main" branch. kwargs (`Dict[str, Any]`, *optional*): Additional keyword arguments passed along to [`~Trainer.create_model_card`]. Returns: The URL of the repository where the model was pushed if `blocking=False`, or a `Future` object tracking the progress of the commit if `blocking=True`. rN/r)r Tr? model_tagsrrr)rr+rtr r,r-rEr)r$rrar^rrrrrr`rCrr3rrrrHrar(rJrrN)rrtrDr rErr model_tags rr_zTrainer.push_to_hubs6ZZ d3  $))"7"7yy%%-!$))"6"67<< !YY3399#>rB * 0C0C    $   E  * t,))+  4::|T 2 >V#!#v&.#."(.!1v!ZZ22 5 F6N26N)))4 5 ?*?? !!#%% ,,)&,!&;%}tA|| | }|shd}tC|dr0tE|jFtHr|jFj4}tA|| |}tA|| |}tA|| |}|jKtC|jLdr9tO|jLjJr|jLjK|jPdk\rd|_)||jT_+tY|D]w\}}|j[||||\}}}t1|jdd}d|j\vr|j_||nd}|.|ja| }| |nt'jb| |fd} || |n te| |d} || |n te| |d} |||n te||d}|jTjg||jh|jj|_5|jjlr{|jno| m| k|jjpjr}i} d|j\vr| nd| d<d|j\vr|nd| d<|jotud(| | d| |}|jjls%|jv|dz|jvzdk(s|jy|j{| d|scjy|j{| d jy|j{| d!jy|j{|d"~ ~ ~ ~t&j|jd#\} } } }z|jPrtC|d$r t|d$|jy|j{| d|scjy|j{| d jy|j{| d!jy|j{|d"|j}!|sjnd}"|sjnd}#|sjnd}$|jnb|"`|#^|jjlsHd|j\vr|!nd|d<d|j\vr|$nd|d<|jotud(|"|#d|}n|i}t|}|!$|!jj||d%<t|jD]0}%|%j|d&r|j|%||d&|%<2t|"|#|| 'S))rz+dataloader must implement a working __len__NrTrFrrr_is_accelerate_preparedz\Batch size cannot be None. Ensure the dataloader has a valid batch_size or total_batch_size.rrrrr )make_multiple_ofrrrrrrh)rrrrrrr eval_losses eval_predseval_label_idseval_inputs_ids)NNNNrrrQrr)Jrr_rrrCr#r"rrr7rOrrJrrrNr$rGrrrr8rHrsr3rrrrrrtrr:r;rrr@rrTrdrrrYrrrrrKrcatrGrrprwrrrrrQr add_arrays_gather_and_numpifyrr`r=finalizer[rrrrrr$rP)&rrrrrrrrQrrr losses_host preds_host labels_host inputs_hostr&rreval_losses_gathererrLpreds_gathererlabels_gathererinputs_gathererrWrhrrrgrrrrr eval_losspredsr inputs_idsrs& rrzTrainer.prediction_loopsyy*%JK K7K7W3]a]v]v  $ $)?!$1MDAq  e  S t'' (A -%4::2E,,0D0D  ((/%%33E43P  ##" DJJ&%*"((!%!3!3""u}}T[[I$$u~~dkkJz#>6 *x8K8K/L NN   ! ??a DJ0:-%j1. [LD&#'#7#7vG[it#7#u D&&%djj2C[QO@HDLdLd@d##F?$;&>z<&X Y  & &t'?'? M]'^ _  & &t'?'? M^'_ `(113 1E'')46JO,,.PT 7K_--/QU   ,!%II005;t?W?W5W ]aOH %6>$BZBZ6Z `dOH %**>+teW`+tds+tuG _G(0  3<>>3C3H3H3JG()/ 0 ' IC>>%6$7q"9:8? C8H,-Qse45 I%9gcopprc|ytrt||}t|Strt|}t|S|jj t jk(r t|}t|S)rN) rrJrrrrDrfrErBrIrs rrSzTrainer._gather_and_numpifysy ?  ! #,Wd;G g&& % & )Gg&&YY $ $ (@(@ @(1Gg&&rc|jsyddg}tjjtjj |j j drWttjj |j j d5}|j}dddnd}}|D]&}||vs|jdr||z }|d|z }(||k7rpttjj |j j dd5}tjd||j|ddd|j jdtjd |j j!s6|j j#d |j j%yy#1swY-xYw#1swYxYw) z8Add SageMaker Checkpointing patterns to .gitignore file.Nz*.sagemaker-uploadingz*.sagemaker-uploadedz .gitignorer rz"Writing .gitignore file. Content: g?z'Add *.sagemaker patterns to .gitignore.)rrbrArrrepo local_dirr r5r[rr/r!git_addrsleep is_repo_clean git_commitgit_push)rpatternsrcurrent_contentcontentpatterns r_add_sm_patterns_to_gitignorez%Trainer._add_sm_patterns_to_gitignores))+ +-CD 77>>"'',,tyy':':LI Jbggll499#6#6 EF +!"#&&( + +!O" .Gg%##D)w&GG9~-G  . o %bggll499#6#6 EsK !q A'KL  ! ,' 3yy&&( II !J K II   )3 + +  ! !s$G#0*G0#G-0G9c x i}tdr@|jjj |jjj}d|vr8|jjdkDr t d|d|j_|jjj }tdrSgd}td i|Dcic]}||j|c}}tdr|jj|_ |jd}tds |r?td |r+|jjstjd |_|jd d |jji}tdr|d <n|j!||jj"dkDred|_t'j(t*t'j(dkDr$t-|jj"|d<n t dt/d i||_|j0j2|_dt7j8|j4j:j=vr:t?j@|j4|jjB|_tE|j0jFd ddu|_$tE|j0jFdddu|_%tE|j0jFdddu|_|jJr|j0jFjL}dD]=}tO|||jjPjS|tE||?|jTr!|jjVr t d|jHr'tE|jdd|jY|jjZrL|jHs |jJr4|jj\r|jHrdnd} t | d|jHrN|j0jFjj^dk(r!|jj`r t d|jjZrN|jJrAdtc|j0jFjLjdvr t dyyycc}w)!Nrrr zThe `AcceleratorConfig`'s `num_steps` is set but `gradient_accumulation_steps` is greater than 1 in the passed `TrainingArguments`If using the passed `AcceleratorConfig` is desired, do not set the `TrainingArguments` `gradient_accumulation_steps`.) split_batchesdispatch_batches even_batchesuse_seedable_samplerz1.1.0 non_blockingrzp`non_blocking` is only supported in accelerate v0.30.0 and above. Please upgrade accelerate to use this feature.zx`non_blocking` is enabled but `dataloader_pin_memory` is not. For the best performance, it's recommended to enable both.gradient_accumulation_kwargsr!dataloader_configTr)tp_sizetorch_tp_pluginz4Requires accelerate>1.3.0 to use Tensor Parallelism.use_gather_object)ryr)limit_all_gathersactivation_checkpointingzThe activation_checkpointing in FSDP config and the gradient_checkpointing in training arg can't be set to True simultaneously. Please use FSDP's activation_checkpointing logic when using FSDP.r DeepSpeedrzJ can't be used with `save_only_model` along with `load_best_model_at_end`.zo`auto_find_batch_size` isn't supported yet with DeepSpeed Zero-3. Please consider using Zero-2, Zero-1, or FSDPSHARDED_STATE_DICTzWsave_only_model option is not compatible with FSDP state dict type 'SHARDED_STATE_DICT'r)3ruraccelerator_configrurrto_dictrr$ data_seedr=rrr<rtr!rbrw is_tp_enabledrrryrrrOrrrrrrrreval_use_gather_objectr3rprCrJrrrArr{rrrr zero_stagerrrx) rgrad_acc_kwargsrrrrvrtrrwrappers rr%z*Trainer.create_accelerator_and_postprocesss "8 ,1M1M1j1j1v"ii::WWO / )yy44q8 L 9H 8T 5!YY99AAC "8 , m  7!EVWE5,0077W! 'w/.2ii.A.A!+)--n= &x0!GDII$C$CO.:  *=>  : :  #8 ,(9D$ % KK* + 99  q !%D }}/07==3II*CDIIL]L]*^&' !WXX'..#//BB '"3"3D4H4H"I"T"T"Y"Y"[ [#,#4#4$$ 8X8X$D %,D,<,<,B,BDVX\$]ei$i!&t'7'7'='=}dS[__$T%5%5%;%;=NPTU]aa   **00<%>kFGy(rst t  % %  &&77BBaG ..B  II % %$$$D,<,<,B,B,N,N,^,^(__vw w`% &iXsT7 cddlm}|jjj}||j j |_|j j |_|j j|j|y)zO Sets values in the deepspeed plugin based on the Trainer args rrN) rrrOrpr!rrdeepspeed_configr r)rrr ds_plugins rrz#Trainer.propagate_args_to_deepspeedYsg Q$$**;; !9):P:P:W:W!X %.%;%;%B%B "55diiAUVrc|jr`t|jrIddlm}ddlm}t|jj|r5||j|jjj_ t|jddtjk(r|jj j"j$j&rt)j*t,t)j*dkDrZ|jjjj/|jj j"j$dyyyyyy)Nr) PeftConfig)fsdp_auto_wrap_policyrz0.27.0T)r)rJrrrrpeft.utils.otherrractive_peft_configrOrprrr3rrKrr?bnb_4bit_quant_storagerMrrryset_mixed_precision)rrrs rrz"Trainer._fsdp_qlora_plugin_updateses    N4::$> ' >$**77DF[\`\f\fFg  &&22C $94@DVDeDeeJJ++??VVhhMM"45 h8OO  &&22FFJJ++??VVaeGPif%? rcg}d}t|D]} |jt|t |dkDxr%d|dvxr|j xs|j du}|r: t|Dcgc]$}|djdj&c}}||jjr)|jj|j}tj |rN|j#|}|jj$dkDr$|j'dk(r|j)d}||fS#t$rY&wxYwcc}w#ttf$rYwxYw)Nrrgrr )rrarT StopIterationr7rQrr+nerrrrjrOrr is_tensorrrur unsqueeze) rr`rdr8rerfrQcount_num_items_in_batchr s rrzTrainer.get_batch_samplesus !{# A $$T.%9:    " M!,,  ..6))5 ! $ %(Xe)fu5?+=+=d+C*H*H*J)f%g"  )yy66%)%5%5%<%<=O%P%T%T%V"12%7%:%:6%B"99??Q&+=+A+A+Cq+H);)E)Ea)H&000I!  **g~.  s5E ( E1)EE EEEE10E1rLc|j}|dk}t|r t|nd}|=t||jzd}|r"t j |j|z}|r{|j|}|jdkDr|zt||zdkDz} ||z} nt j |j} |j||jz} nX|jdkDr1tj} |}||jz}|j|z} ntd|j| || |||fS)a Calculates and returns the following values: - `num_train_epochs` - `num_update_steps_per_epoch` - `num_examples` - `num_train_samples` - `epoch_based` - `len_dataloader` - `max_steps` rNr zYargs.max_steps must be set to a positive value if dataloader does not have a length, was ) rer_r7rtrrrrfrrsysmaxsizer) rrrrLrerOrPrMrrfrNs rrz#Trainer.set_initial_training_valuessiNN !m ,6z,BZ  %),^t?_?_-_ab)c & IId&;&;>X&XY  ,,Z8L~~!#,0J#JS ::Q>N$ %.0F$F!#'99T-B-B#C $($5$5j$ADDYDY$Y! ^^a "{{ )2 &1DNNBL $1G G NN#%   &       r) NNNNNNNNNNrNN)rNr)F)TN)NNN)NNNNN)NNminimizeNNr)FN)NF)NNr)Nr ) NNNNNNNNN)zEnd of trainingTNN)r2 __module__ __qualname____doc__trainer_pt_utilsrrrrrrr r'rModulergrrrrrrr1r!rr.rrQrr7rrr& OptimizerrULambdaLRrrrrpropertyrsetterrrrZrrrLrrrrgrhSamplerrrrrrr rrrrr<rB parameterrrD staticmethodr#rrr r%rr=r7r,rfrrorrrr8rrDrrArrr4rrrrrrrr r rRrOrErCrKrVrTrSr%r]rrrCrurvrBrNrErrrTrrPrrrr)r`r(rrJr_rrSrnr%rrrrrrrrr<s \~lk[+=wdhi:>"&04W[Y] >B04FJ59jvaehl!Mc_bii56Mc Mc - Mc  g@R&R ST Mc uWd3<.@BT%TUV Mc# )+=?UWee f McXb/&9:;Mc$H-Mc"(N+;T+A"BCMcD12Mc(5;;#8#898EKKD\D\DeDe;ffgMc#+5ekk6K6K1LdSVX[S[n1\+]"^Mc (0%,, 9UW\WcWc9c0d'e!McjMc^ %8$;<%%11 &<" 5< 8 \;.@;xX[};BEI'%'4B=A48 o (sDy)9 :o^T#s(^T9:o'tCy1 odimcHJ(V=pxt ,hl'^R)_V$"L2Jh0id?1Bk5Z1(52%9RLPKO+5;?=AR8^$4d3:6F$FGHR$Hd3:.>-?-F$GHR R d3i( R % 678 R(N#3S#89:R wW % &Rh_S%Z(_huo_QU_.5s):#;ellTWFW@X$d3ellC6G0H+H&IdSVX]^c^j^jlo^oXpSpNq 5 HTN aeN!YYN!(,S% c8I2J-J(KN! N!`2;h2t2 0t 01:Xc]1:41:fF>HSMF>P)L )LV "1FRW" c"6:6FJ+/!' fuWd3<.@%@ABfd3i(f f c5j  fRhn>t#>t2:492E>tad>t >tH04+/!' TuTuTu'tn Tu d3i( Tu  Tu Tul0,0 g&yyg&S% c 1223g&# g& d3i( g& x %x '=x ?UU V g&Rc5s9J3K.K)L$%(3-%(#'!%,0$((,-148/348Hh3-Hh#HhCcD() Hh SM Hh ! HhS$s)T)*HhCcD01HhsDIt+,HhCcD01HhT@9D4):#"& J   J J } J 3- J  J h04+/!' xqxqxq'tn xq d3i( xq  xq xqt' %!Noxb W +1Z: %: 3=: WZ: rr(rrr,rr$importlib.metadatarrr3rrbrrrHrr5rr.collections.abcrpathlibrtypingrrrrr integrationsr huggingface_hub.utilsrgrrrrtorch.distributed distributedr@huggingface_hubr r r packagingrrtorch.utils.datarrrrrrrconfiguration_utilsrdata.data_collatorrrr debug_utilsrr!feature_extraction_sequence_utilsrfeature_extraction_utilsrrErr image_processing_utilsr!integrations.deepspeedr"r#r$integrations.tpur% modelcardr&modeling_utilsr'r(r)models.auto.modeling_autor*r+ optimizationr,r-processing_utilsr. pytorch_utilsr/r0tokenization_utils_baser1trainer_callbackr2r3r4r5r6r7r8r9rr:r;r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrLrM trainer_utilsrNrOrPrQrRrSrTrUrVrWrXrYrZr[r\r]r^r_r`rarbrcrdrwrerfrgrhrirjrkrlrmrnrorprqrrrsrtrurvrwrxryrzr{r|r}r~rrrrrrrrrrrrrrrrrrutils.deprecationrutils.quantization_configrrWr\utils.notebookrr rrtorch_xla.core.xla_modelr xla_modelrtorch_xla.debug.metricsr/r&r; torch_xla XLA_VERSIONrrtorch_xla.distributed.spmdspmdrtorch_xla.runtimeruntimer!smdistributed.modelparallel.torch modelparallelrosmdistributed.modelparallel SMP_VERSIONrnrrrrsafetensors.torchrrr acceleraterrryaccelerate.staterrrrrrrrr DATA_SAMPLERSraccelerate.data_loaderrrrrrrr. get_loggerr2rrrErHr-rrIr|rrrrrs    #@@  - AAcc1\\<G<`6ee1&RR3,=   ,2KJ***********V/9)),8 8))))4*W]];7=7==I_;``//&"33F - k :mgmmF>S Sgg %:<1#OMw}}'(=7==+AA>w}}'(=7==+BB@/00 >8$8 7&   H %)) $&`N `N r