o }oi**@sddlmZddlmZddlZddlZddlm Z m Z ddl m Z ddl mZddlmZddlmZmZmZddlmZmZdd lmZdd lmZdd lmZdd lmZd Z ej!j"e ddej#ej$fddZ%ej!j"e e dddddej&ddddddddddddddddd dd!d"de fd#ee'd$e'd%e(d&e(d'eej)d(ee(d)e(d*e*d+e(d,e(d-e(d.e'd/e(d0e+d1e(d2e(d3e(d4e(d5e*dej,f(d6d7Z-d8ej,dej,fd9d:Z.ej!j"e e d     ; d@d#ee'd$e'd+e(d,e(dd?Z/dS)A)OptionalN)finetunepretrain)MockDataModule) PEFT_STR2CLS)default_finetune_recipe) default_logdefault_resumetensorboard_logger)nemotron_modelnemotron_trainer),distributed_fused_adam_with_cosine_annealing)GarbageCollectionCallback)MegatronCommOverlapCallback)TimingCallback nemotron3_22bnamereturncCs ttdS)aj Factory function to create a Nemotron3 22B model configuration. Returns: run.Config[pl.LightningModule]: Configuration for the Nemotron3 22b model. Examples: CLI usage: $ nemo llm pretrain model=nemotron3_22b ... Python API usage: >>> model_config = model() >>> print(model_config) )version)r NAMErr^/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/llm/recipes/nemotron3_22b.pymodel#s r)targetrdefault Fiz bf16-mixedg? iiigh㈵>-C6?dirrtensor_parallelismpipeline_parallelismpipeline_parallelism_typevirtual_pipeline_parallelismcontext_parallelismsequence_parallelism num_nodesnum_gpus_per_node max_steps precisionaccumulate_grad_batchesgradient_clip_vallimit_test_batcheslimit_val_batcheslog_every_n_stepsval_check_intervalperformance_modecCstj|ttdid|d|d|d|d|d|d|d| d | d | d | d |d |d|d|dttgtjt|||dt||t|ddt | ||||| dt d}|rbt |}|S)a Create a pre-training recipe for Nemotron3 22B model. This function sets up a complete configuration for pre-training, including model, trainer, data, logging, optimization, and resumption settings. Args: dir (Optional[str]): Directory for saving logs and checkpoints. name (str): Name of the pre-training run. 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. gradient_clip_val (float): Value for gradient clipping. 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. global_batch_size (int): Global batch size. micro_batch_size (int): Micro batch size. seq_length (int): Sequence length. warmup_steps (int): Number of warmup steps. constant_steps (int): Number of constant steps. min_lr (float): Minimum learning rate. max_lr (float): Maximum learning rate. performance_mode (bool): If true, enables optimizations for maximum performance. fn (Callable): The pre-training function to use. Returns: run.Partial: Partial configuration for pre-training. Examples: CLI usage: $ nemo llm pretrain --factory nemotron3_22b $ nemo llm pretrain --factory "nemotron3_22b(num_nodes=1, name='my_nemotron_pretrain')" Python API usage: >>> recipe = pretrain_recipe(name="nemotron_pretrain", num_nodes=1) >>> print(recipe) Note: This recipe uses a mock dataset, look for the finetune examples to see how to change the dataset. r$r%r&r'r(r)r*r+r,r-r.r0r1r2r3 callbacks) seq_lengthglobal_batch_sizemicro_batch_sizer)r#rr )r- warmup_stepsconstant_stepsmin_lrmax_lr clip_grad)rtrainerdatalogoptimresumeNr) runPartialrr Configrrrr r r "pretrain_performance_optimizations)r#rr$r%r&r'r(r)r*r+r,r-r.r/r0r1r2r3r7r8r6r9r:r;r<r4fnreciperrrpretrain_recipe7spU      'rIrHcCsT|jjsg|j_tjtddd}tjtddddd}|jj||gd|jj_|S)a Create a performance-optimized pre-training recipe for Nemotron3 22B model. This method enables performance optimizations that may not be suitable for all use cases. It builds upon the standard pre-training recipe and adds additional performance enhancements. Args: recipe (run.Partial): Base pre-train recipe to which performance optimizations will be added Returns: run.Partial: Partial configuration for performance-optimized pre-training. Note: Use this method with caution and only when you need maximum performance. It may not be suitable for all hardware configurations or use cases. d)gc_interval_traingc_interval_valTF)tp_comm_overlapdefer_embedding_wgrad_computewgrad_deferral_limit(overlap_param_gather_with_optimizer_step) r>r5rCrErrextendpluginsgrad_reduce_in_fp32)rHgarbage_collection_callbackmcomm_overlap_callbackrrrrFs* rFlora peft_schemepacked_sequencecCsttd|||||}|dus|dkr d|jj_d|jj_n|dvr6t t ||_ d|jj_nt d|d |jj_|S) a Create a fine-tuning recipe for Nemotron3 22B model. This function sets up a complete configuration for fine-tuning, including model, trainer, data, logging, optimization, and resumption settings. The recipe uses LoRA (Low-Rank Adaptation) for efficient fine-tuning, unless peft_scheme is set to None. Args: dir (Optional[str]): Directory for saving logs and checkpoints. name (str): Name of the fine-tuning run. num_nodes (int): Number of compute nodes to use. num_gpus_per_node (int): Number of GPUs per node. peft_scheme (Optional[str]): Name of the peft scheme to use for fine-tuning. Allowed values: 'lora'/'dora'/'none'/None. packed_sequence (Optional[bool]): Packing multiple training sequences into one long sequence for training efficiency. Default sequence length is 2048. Returns: run.Partial: Partial configuration for fine-tuning. Examples: CLI usage: $ nemo llm finetune --factory nemotron3_22b Python API usage: >>> recipe = finetune_recipe(name="nemotron3_22b_finetune", num_nodes=8) >>> print(recipe) Note: This recipe uses the SQuAD dataset for fine-tuning. zthhaus/nemotron3-22b-hfNnoner gh㈵>)rWdorar"zUnrecognized peft scheme: F)rrlowerr>strategytensor_model_parallel_sizerAconfiglrrCrErpeft ValueErrorcross_entropy_loss_fusion)r#rr*r+rXrYrHrrrfinetune_recipes)     rd)Nrrr rWF)0typingrlightning.pytorchpytorchplnemo_runrCtorchnemo.collections.llm.apirr"nemo.collections.llm.gpt.data.mockrnemo.collections.llm.peftr-nemo.collections.llm.recipes.finetune_defaultr(nemo.collections.llm.recipes.log.defaultrr r %nemo.collections.llm.recipes.nemotronr r 'nemo.collections.llm.recipes.optim.adamr 3nemo.lightning.pytorch.callbacks.garbage_collectionr6nemo.lightning.pytorch.callbacks.megatron_comm_overlaprnemo.utils.exp_managerrrclifactoryrELightningModulerbfloat16strintdtypeboolfloatrDrIrFrdrrrrs               ! .