U 9%e3@sdZddlZddlZddlmZddlmZmZddlm Z ddl m Z m Z ddl mZmZmZd d lmZmZmZmZmZmZerddlZdd lmZd d lmZerddlmZ e!e"Z#e$e$d ddZ%e&e&e&dgdddZ'GdddeZ(dS)zD Benchmarking the library on inference and training in PyTorch. N)wraps)CallableOptional)PretrainedConfig)TF_MODEL_MAPPINGTF_MODEL_WITH_LM_HEAD_MAPPING)is_py3nvml_availableis_tf_availablelogging) BenchmarkMemory MemorySummarymeasure_peak_memory_cpustart_memory_tracingstop_memory_tracing)ResourceExhaustedError)TensorFlowBenchmarkArguments do_eager_modeuse_xlacsfdd}|S)NcsXtfdd}ttjdfdd}dkrPdk rLtd|S|SdS) Ncs ||SNargskwargsfuncrb/var/www/html/Darija-Ai-API/env/lib/python3.8/site-packages/transformers/benchmark/benchmark_tf.pyrun_in_eager_mode5szFrun_with_tf_optimizations..run_func..run_in_eager_mode)Zexperimental_compilecs ||Srrrrrrrun_in_graph_mode9szFrun_with_tf_optimizations..run_func..run_in_graph_modeTFzcCannot run model in XLA, if `args.eager_mode` is set to `True`. Please set `args.eager_mode=False`.)rtffunction ValueError)rr r!rrrrun_func4s z+run_with_tf_optimizations..run_funcr)rrr%rrrrun_with_tf_optimizations3sr&z tf.Tensor) batch_sizesequence_length vocab_sizereturncs:tfddt||D}tj|||ftjdS)Ncsg|]}ddqS)rr )randint).0irngr)rr Lsz$random_input_ids..)shapeZdtype)randomRandomranger"ZconstantZint32)r'r(r)valuesrr.rrandom_input_idsJsr6c@seZdZUeed<eed<dZeed<eddZ ee e e ddd Z ee e e dd d Z ee e eeegdd d Zee e eeegdddZee e egdfdddZee e egdfdddZe dddZegdfeegdddZdS)TensorFlowBenchmarkrZconfigsZ TensorFlow frameworkcCstjSr)r" __version__)selfrrrframework_versionUsz%TensorFlowBenchmark.framework_version) model_namer'r(r*cCs0|jj}|dkrtd||||}||SN@A device strategy has to be initialized before using TensorFlow.)rstrategyr$_prepare_inference_func_measure_speedr:r<r'r(r? _inferencerrr_inference_speedYs z$TensorFlowBenchmark._inference_speedcCs0|jj}|dkrtd||||}||Sr=)rr?r$_prepare_train_funcrAr:r<r'r(r?_trainrrr _train_speedas z TensorFlowBenchmark._train_speedcCsT|jjr$tjj|jj|jjd|jj}|dkr) ris_gpur"config experimentalset_memory_growthgpu_list device_idxr?r$r@_measure_memoryrBrrr_inference_memoryhsz%TensorFlowBenchmark._inference_memorycCsT|jjr$tjj|jj|jjd|jj}|dkr.encoder_decoder_forwardcs ddS)NF)r[rrr\rrencoder_forwardszDTensorFlowBenchmark._prepare_inference_func..encoder_forward) config_dictrfp16NotImplementedErrorhasattr isinstancerTlistlenonly_pretrain_model __import__getattr ImportErrorr __class__r)encoderr6r& eager_moderis_encoder_decoder) r:r<r'r(rKhas_model_class_in_config model_classtransformers_module model_clsr)r^r_rCrr\rr@s6         z+TensorFlowBenchmark._prepare_inference_funcc s<|j|}|jjdk rtd|jjr.tdt|doPt|jt oPt |jdk}|jj s|rz2d|jd}t d|gd}t ||}||Wqtk rt|d YqXnt|j|t|d r|jn|jj} t||| t|jj|jjfd d } t|jj|jjfd d} |jr4| n| } | S)NFzWTraining cannot be done in eager mode. Please make sure that `args.eager_mode = False`.rSrTrrUrVrWrYr)cs&ddd}t|j}|S)NT)rZlabelsr[rr" gradientsZtrainable_variablesZlossrur\rrencoder_decoder_trainszFTensorFlowBenchmark._prepare_train_func..encoder_decoder_traincs$ddd}t|j}|S)NT)rsr[rrtrvr\rr encoder_trainsz>TensorFlowBenchmark._prepare_train_func..encoder_train)r`rrmr$rarbrcrdrTrerfrgrhrirjrrkr)rlr6r&rrn) r:r<r'r(rKrorprqrrr)rwrxrGrr\rrEs:          z'TensorFlowBenchmark._prepare_train_func)r*c Cs|jjzX|jjs |jjr:tdtj|dddtj||jjdd}t |dWW5QRSt k r}z| d|W5d}~XYnXW5QRXdS)NzCDo inference on TPU. Running model 5 times to stabilize compilationr )repeatnumber g$@Doesn't fit on GPU. ) rr?scopeis_tpurloggerinfotimeitrzminrprint_fn)r:rZruntimeserrrrAs z"TensorFlowBenchmark._measure_speed)rr*c Cs|td|jjZz |jjr>|jjs6tdtd}|jj rPt dn|jj rt snt dd}qtdt|t|jj}t|}|j}t|}tn6|jjrtdd}nt|}t|trt|n|}|jjrt|}|dkr|j}nd}||fWW5QRStk rl} z$|d | WYW5QRd Sd} ~ XYnXW5QRXdS) NzNote that TensorFlow allocates more memory than it might need to speed up computation. The memory reported here corresponds to the memory reported by `nvidia-smi`, which can vary depending on total available memory on the GPU that is used.zu`args.eager_mode` is set to `False`. Make sure to run model in eager mode to measure memory consumption line by line.rVzuMemory Benchmarking is currently not implemented for TPU. Please disable memory benchmarking with `args.memory=False`zypy3nvml not installed, we won't log GPU memory usage. Install py3nvml (pip install py3nvml) to log information about GPU.N/AzlMeasuring total GPU usage on GPU device. Make sure to not have additional processes running on the same GPU.z\When enabling line by line tracing, the max peak memory for CPU is inaccurate in TensorFlow.r})rN)rrrr?r~Ztrace_memory_line_by_linermr$rrrbrJr warningnvmlZnvmlInitZnvmlDeviceGetHandleByIndexrOZnvmlDeviceGetMemoryInfousedrZ nvmlShutdownrrdintrtotalrr) r:rtraceZmemoryhandleZmeminfoZmax_bytes_in_useZ memory_bytessummaryrrrrrPs^    z#TensorFlowBenchmark._measure_memory)__name__ __module__ __qualname__r__annotations__rr8strpropertyr;rfloatrDrHrrrrQrRrr@rErArPrrrrr7Ps*     )0r7))__doc__r2r functoolsrtypingrrZconfiguration_utilsrZmodels.auto.modeling_tf_autorrutilsr r r Zbenchmark_utilsr rrrrrZ tensorflowr"Z'tensorflow.python.framework.errors_implrZbenchmark_args_tfrZpy3nvml.py3nvmlZpy3nvmlrZ get_loggerrrboolr&rr6r7rrrrs$