o }oi$@s,ddlmZddlmZddlZddlZddlm Z ddl m Z ddl m ZddlmZmZmZmZmZddlmZmZd)d ed ed ejejfd d Z                d*dededeejdeedededededed ed!ed"ed#ed$ed%ed&eeeje d ejej f"d'd(Z!dS)+)OptionalN)Callback)DistributedDataParallelConfig) lightning) BertModelHuggingFaceBertBaseConfigHuggingFaceBertLargeConfigMegatronBertBaseConfigMegatronBertLargeConfig) bf16_mixed fp16_mixedmegatronversion bert_typereturncCsd}|dkr|dkrtt}n)|dkr|dkrtt}n|dkr,|dkr,tt}n |dkr9|dkr9tt}|dusGJd|d|tjt|dS) aC A function to create a Bert models. Args: version (str): The version of the Nemotron model to create. one of ["bert_110m", "bert_340m"]. bert_type (str): The Bert type. either "megatron" or "huggingface". Returns: run.Config[pl.LightningModule]: Configuration for the Bert model. N bert_110mr huggingface bert_340mzInvalid BERT version: z | )config)runConfigr rr rr)rrrrU/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/llm/recipes/bert.py bert_model"s     rF{ bf16-mixed tensor_parallelismpipeline_parallelismpipeline_parallelism_typevirtual_pipeline_parallelismcontext_parallelismsequence_parallelism num_nodesnum_gpus_per_node max_steps precisionaccumulate_grad_batcheslimit_test_batcheslimit_val_batcheslog_every_n_stepsval_check_interval callbackscCstjtj||||||ddddtjtddddddd }d}| dkr%t}n| dkr,t}tjtjd||| | | | ||||d|d }|S) a Configure the NeMo Lightning Trainer for BERT models. This function sets up the distributed training strategy and other training parameters. Args: tensor_parallelism (int): Degree of tensor model parallelism. pipeline_parallelism (int): Degree of pipeline model parallelism. pipeline_parallelism_type (Optional[torch.dtype]): Data type for pipeline parallelism. virtual_pipeline_parallelism (Optional[int]): Size of virtual pipeline parallelism. context_parallelism (int): Degree of context parallelism. sequence_parallelism (bool): Whether to use sequence parallelism. num_nodes (int): Number of compute nodes to use. num_gpus_per_node (int): Number of GPUs per node. max_steps (int): Maximum number of training steps. precision (str): Precision configuration, one of fp32, 16-mixed or bf16-mixed. accumulate_grad_batches (int): Number of steps per gradient accumulation. limit_test_batches (int): Limit the number of test batches. limit_val_batches (int): Limit the number of validation batches. log_every_n_steps (int): Log every n steps. val_check_interval (int): Run validation every N steps. callbacks (Optional[list[run.Config[Callback]]]): List of callback configurations. Returns: run.Config[nl.Trainer]: Configuration for the NeMo Lightning Trainer. TF)check_for_nan_in_gradgrad_reduce_in_fp32overlap_grad_reduceoverlap_param_gatheraverage_in_collective) tensor_model_parallel_sizepipeline_model_parallel_sizepipeline_dtype$virtual_pipeline_model_parallel_sizecontext_parallel_sizesequence_parallelgradient_as_bucket_viewckpt_include_optimizerckpt_async_saveckpt_parallel_loadddpNz16-mixedrgpu) acceleratorr1devicesr,r-r.r/r*r(pluginsstrategyuse_distributed_samplerr0)rrnlMegatronStrategyrr r Trainer)r"r#r$r%r&r'r(r)r*r+r,r-r.r/r0r1rFprecision_plugintrainerrrr bert_trainer;sV,rM)r )rrNNrFrrrrrrrr r!N)"typingrlightning.pytorchpytorchplnemo_runrtorch$lightning.pytorch.callbacks.callbackrmegatron.core.distributedrnemorrHnemo.collections.llmrrrr r 6nemo.collections.llm.recipes.precision.mixed_precisionr r strrLightningModulerintdtypeboollistrJrMrrrrsz