o wi+$@sddlmZddlmZddlZddlZddlm Z ddl m Z ddl m ZddlmZmZmZddlmZmZded ejejfd d Z           d'dededeejdeedededededededed ed!ed"ed#ed$eeeje d ejejf"d%d&ZdS)()OptionalN)Callback)DistributedDataParallelConfig) lightning)BertEmbeddingLargeConfigBertEmbeddingMiniConfigBertEmbeddingModel) bf16_mixed fp16_mixedversionreturncCsNd}d|vr tt}n d|vrtt}|dus Jd|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. N340m110mzInvalid BERT version: )config)runConfigrrr)r rrh/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/nemo/collections/llm/recipes/bert_embedding.pybert_embedding_models   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$virtual_pipeline_model_parallel_sizecontext_parallel_sizesequence_parallelgradient_as_bucket_viewckpt_include_optimizerckpt_async_saveckpt_parallel_loadddpNz16-mixedrgpu) acceleratorr,devicesr'r(r)r*r%r#pluginsstrategyuse_distributed_samplerr+)rrnlMegatronStrategyrr r Trainer)rrrr r!r"r#r$r%r&r'r(r)r*r+r,rAprecision_plugintrainerrrr bert_trainer1sV,rH)rrNNrFrrrrrrrrrN) typingrlightning.pytorchpytorchplnemo_runrtorch$lightning.pytorch.callbacks.callbackrmegatron.core.distributedrnemorrCnemo.collections.llmrrr6nemo.collections.llm.recipes.precision.mixed_precisionr r strrLightningModulerintdtypeboollistrErHrrrrsz