o }oi$@@s~ddlmZmZddlmZddlZddlZddl m Z ddl m Z ddl mZddlmZmZddlmZddlmZdd lmZmZdd lmZdd lmZdd lmZm Z m!Z!dd l"m#Z#ddl$m%Z%ddl&m'Z'ddl(m)Z)ddl*m+Z+ddl,m-Z-dZ.ej/j0e.ddej1ej2fddZ3ddej4dddddddddf de5de5d eej6d!ee5d"e5d#e7d$e7d%e7d&e5d'e5d(e5d)ee8ej1e dej1ej9fd*d+Z:ej/j0ee.d,dd-d.dd/efd0ee;d1e;d&e5d'e5d2e7d3edejej/j0ee.d,  - 9  :   /dBd0ee;d1e;d&e5d'e5d;ee;dd?Z?d6ej>> model_config = model() >>> print(model_config) )config)runConfigr seq_lengthr )confr"]/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/llm/recipes/llama31_405b.pymodel,s r$T@i{tensor_parallelismpipeline_parallelismpipeline_parallelism_typevirtual_pipeline_parallelismcontext_parallelismsequence_parallelism'account_for_embedding_in_pipeline_split"account_for_loss_in_pipeline_split num_nodesnum_gpus_per_node max_steps callbacksc Csftjtj||||||||dddtjtddddddd } tjtjdd| | ddd| |t| d d d } | S) aC Configure the NeMo Lightning Trainer for Llama3.1 405B model. This function sets up the distributed training strategy optimized for the large 405B model. 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. callbacks (Optional[list[run.Config[Callback]]]): List of callback configurations. Returns: run.Config[nl.Trainer]: Configuration for the NeMo Lightning Trainer. Examples: CLI usage: $ nemo llm pretrain trainer=llama31_405b ... Python API usage: >>> trainer_config = trainer(num_nodes=4, num_gpus_per_node=8) >>> print(trainer_config) Note: This configuration uses extensive parallelism to handle the large model size efficiently. Tcheck_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_parallelr/r0gradient_as_bucket_viewckpt_async_saveckpt_parallel_loadddpgpu2 Fi) acceleratoraccumulate_grad_batchesr4deviceslimit_test_batcheslimit_val_batcheslog_every_n_stepsr3r1pluginsstrategyuse_distributed_samplerval_check_interval)rrnlMegatronStrategyrTrainerr)r)r*r+r,r-r.r/r0r1r2r3r4rQtrainerr"r"r#rWAsN-rW)targetrdefaultrFFdirrperformance_modefnc Csbtj|tt||ttgdtjtddddt||t|ddt dd t d }|r/t |}|S) aj Create a pre-training recipe for Llama3.1 405B 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. num_nodes (int): Number of compute nodes to use. num_gpus_per_node (int): Number of GPUs per node. 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 llama31_405b $ nemo llm pretrain --factory "llama31_405b(num_nodes=4, name='my_405b_pretrain')" Python API usage: >>> recipe = pretrain_recipe(name="llama31_405b_pretrain", num_nodes=4) >>> print(recipe) Note: This recipe is optimized for the large 405B model and requires significant computational resources. )r1r2r4rirF)r global_batch_sizemicro_batch_sizer)rZrrga2U0*3?)max_lr)r$rWdatalogoptimresume) rPartialr$rWrrr rrrr"pretrain_performance_optimizations)rZrr1r2r[r\reciper"r"r#pretrain_recipes & rgrfcCs`|jjsg|j_tjtddd}tjtdtdddd}|jj||gd|jj_ d|j j _ |S)a Create a performance-optimized pre-training recipe for Llama3.1 405B 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_valTrGF)tp_comm_overlaptp_comm_overlap_cfgdefer_embedding_wgrad_computewgrad_deferral_limit(overlap_param_gather_with_optimizer_step) rWr4rrrrrextendrPr7rbruse_precision_aware_optimizer)rfgarbage_collection_callbackmcomm_overlap_callbackr"r"r#res.   relora peft_schemer packed_sequencec Csz|dur|}|dur d}|dur%|dus|dkrd}n|dvr%d}ttd|||||}|dus:|dkrNd|jj_d |jj_d |j_d |j j _ nB|dvrt t||_d |j_d |j_d|j j _d|jj _d|jj_d |jj_d|jj_d |j_d|j j _ ntd|||jj _||j_|rddi|j_t j t|d|j_|rt||}d|jj_d|jj_|S)ax Create a fine-tuning recipe for Llama3.1 405B 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. seq_length (int): Maximum number of tokens per microbatch. packed_sequence (Optional[bool]): If true, fine-tuning sequences will be packed into batches up to the given maximum seq_length for better efficiency. By default, this value equals performance_mode. performance_mode (bool): If true, enables optimizations for maximum performance. Returns: run.Partial: Partial configuration for fine-tuning. Examples: CLI usage: $ nemo llm finetune --factory llama31_405b $ nemo llm finetune --factory "llama31_405b(num_nodes=3, name='my_llama31_405b_finetune')" Python API usage: >>> recipe = finetune_recipe(name="llama31_405b_finetune", num_nodes=3) >>> print(recipe) Note: This recipe uses the SQuAD dataset for fine-tuning. Be aware that fine-tuning a 405B model requires substantial computational resources. Ninone )rudorartzmeta-llama/Llama-3.1-405Br%gh㈵>rHFr'g-C6?zUnrecognized peft scheme: pad_to_max_lengthT)packed_sequence_size)lowerrr$rWrQr;r<r`r]rbrlrrrr peftdimalphause_distributed_optimizercross_entropy_loss_fusionr> ValueErrorr dataset_kwargsr packed_sequence_specs"finetune_performance_optimizationsr/r0) rZrr1r2rvr rwr[rfr"r"r#finetune_recipesN-                rcCst|jdr |jjdurg|j_|dus|dkrAd|jj_d|jj_tjt dddddd|jj_ |jj tjt ddd d n d |jj_d |jj_d |jj_ d g|j_|jj tjt dd d|jj_d|jj_|jj tt|jj ttddd|jj_|S)a Modify the given recipe to optimize settings for performance. This method enables performance optimizations that may not be suitable for all use cases. Intended to build upon the standard fine-tuning recipe. Args: recipe (run.Partial): Base fine-tuning recipe to which performance optimizations will be added peft_scheme (Optional[str]): Name of the peft scheme to use for fine-tuning. Allowed values: 'lora'/'dora'/'none'/None. Returns: run.Partial: Partial configuration for performance-optimized fine-tuning. Note: Use this method with caution and only when you need maximum performance. It may not be suitable for all hardware configurations or use cases. r4Nrxr%r{TFr5)rkrmrnr' linear_qkv)rkrh)hasattrrWr4rrQr;r<rrrrDappendrr>rtarget_modulesrPr7r@rrrbrrq)rfrvr"r"r#rdsV         r)NrYrtr%ruNNF)Atypingrrlightning.pytorchpytorchplnemo_runrtorch$lightning.pytorch.callbacks.callbackrmegatron.core.distributedrnemorrTnemo.collections.llm.apirr"nemo.collections.llm.gpt.data.mockr -nemo.collections.llm.gpt.data.packed_sequencer $nemo.collections.llm.gpt.model.llamar r nemo.collections.llm.peftr -nemo.collections.llm.recipes.finetune_defaultr(nemo.collections.llm.recipes.log.defaultrrr'nemo.collections.llm.recipes.optim.adamr6nemo.collections.llm.recipes.precision.mixed_precisionr;nemo.collections.llm.recipes.tp_overlap_configs.userbuffersr nemo.lightning.pytorch.callbacksr6nemo.lightning.pytorch.callbacks.megatron_comm_overlaprnemo.utils.exp_managerrNAMEclifactoryrLightningModuler$bfloat16intdtypeboollistrVrWstrrdrgrerrr"r"r"r#s                    X 90 `