# Copyright (c) Meta Platforms, Inc. and affiliates. # All rights reserved. # # This source code is licensed under the BSD-style license found in the # LICENSE file in the root directory of this source tree. import sys import time from functools import partial from typing import Any, Dict, List, Optional, Tuple from warnings import warn import torch from omegaconf import DictConfig, ListConfig from torch import nn from torch.distributed import destroy_process_group, init_process_group from torch.optim import Optimizer from torch.utils.data import DataLoader, DistributedSampler from torchtune import config, modules, rlhf, training, utils from torchtune.data import CROSS_ENTROPY_IGNORE_IDX, padded_collate_dpo from torchtune.datasets import ConcatDataset from torchtune.modules.peft import ( disable_adapter, DoRALinear, get_adapter_params, get_adapter_state_dict, get_lora_module_names, get_merged_lora_ckpt, LoRALinear, set_trainable_params, validate_missing_and_unexpected_for_lora, ) from torchtune.recipe_interfaces import FTRecipeInterface from tqdm import tqdm log = utils.get_logger("DEBUG") class LoRADPORecipeDistributed(FTRecipeInterface): """ Distributed LoRA DPO recipe for dense transformer-based LLMs such as Llama2. This recipe supports distributed training and can be run on a single node (1 to 8 GPUs). This is based on HF's DPOTrainer in the TRL library: https://github.com/huggingface/trl/blob/main/trl/trainer/dpo_trainer.py#L65 Features: - FSDP. Supported using PyTorch's FSDP APIs. CPU offload of parameters, gradients, and optimizer states is supported via ``fsdp_cpu_offload``. Resharding of parameters after the forward pass is done by default (corresponding to FULL_SHARD sharding strategy), but can be disabled by setting the config ``fsdp_reshard_after_forward`` to False (this corresponds to SHARD_GRAD_OP sharding strategy). DDP is currently not supported. Training on CPU is not supported. - Activation Checkpointing. This can be controlled using the ``activation_checkpointing`` flag. Activation checkpointing helps reduce the memory footprint since we no longer keep activations in memory and instead recompute them during the backward pass. This is especially helpful for larger batch sizes when you're memory constrained. But these savings in memory come at the cost of training performance. In most cases training can slow-down quite a bit as a result of this activation recomputation. - Activation Offloading. This can be controlled using the ``enable_activation_offloading`` flag. Activation offloading is a technique similar to activations checkpointing that helps reduce the memory footprint to prevent OOMs on CUDA and enable bigger batches. Where activations checkpointing drops the activation in the forward to recompute it later in the backward, activations offloading will drop the activation in the forward to the CPU and bring it back during the backward pass. As always, there is a tradeoff--these savings in memory can come at the cost of training performance and CPU resources. To recover some runtime cost, we've added an option to enable offloading on a different stream to permit overlapping with the computation. This option is currently only available on PyTorch 2.5 or later and will be enabled by default if an acceptable torch version is found. Activation offloading can be used in conjunction with activation checkpointing. - Precision. Full fp32 and bf16 training are supported. Precision is controlled using the ``dtype`` flag. When ``dtype=bf16``, all activations, gradients and optimizer states are in bfloat16. In most cases this should halve the memory footprint of full precision (fp32) training, without loss in model quality (will depend on the model, training data and other settings). For GPUs which do not support bfloat16, we fall back to fp32. Mixed precision training and fp16 precision are currently not supported. - Gradient Accumulation. You can simulate larger batch sizes by accumulating gradients. This is controlled using the ``gradient_accumulation_steps`` flag. Total Batch Size = batch_size * number of GPUs * gradient accumulation steps. For example: with batch_size=1, nproc_per_node=2 and gradient_accumulation_steps=32 we get a total batch size of 64. Gradient accumulation is especially useful when you are memory constrained. In this case, accumulating gradients might give you better training speed than enabling activation checkpointing. - Checkpointing. Model weights are checkpointed both at the end of each epoch and at the end of training. Currently we checkpoint both the adapter weights (trainable params only) and the complete merged weights (adapter weights added back to the base model). For more details please take a look at our LoRA tutorial (https://pytorch.org/torchtune/main/tutorials/lora_finetune.html). Optimizer State and recipe state (seed, total_epochs, number of epochs run etc) are only saved at the end of a given epoch and used in case of resuming training. Resuming training is controlled by the ``resume_from_checkpoint`` flag. Mid-epoch checkpointing is currently not supported. For more details on the checkpointer, please take a look at our checkpointer deepdive (https://pytorch.org/torchtune/main/tutorials/checkpointer.html). - Logging. Terminal, Disk, WandB and TensorBoard are all supported. The following losses are supported in this recipe: - :class:`~torchtune.rlhf.loss.DPOLoss`: Direct Preference Optimization (DPO). - :class:`~torchtune.rlhf.loss.RSOPLoss`: Rejection Sampling Optimization (RSO). For a full list of example configs for this recipe, run ``tune ls`` on the command line. Each config has example commands for how to kick-off training. Args: cfg (DictConfig): OmegaConf object parsed from yaml file Raises: ValueError: If ``dtype`` is set to fp16. ValueError: If world_size is 1 RuntimeError: If ``dtype`` is set to bf16 and the hardware does not support bf16. RuntimeError: If ``enable_activation_offloading`` is True and device is not CUDA. RuntimeError: If ``enable_activation_offloading`` is True and ``enable_activation_checkpointing`` is False. """ def __init__(self, cfg: DictConfig) -> None: self._device = utils.get_device(device=cfg.device) self._dtype = training.get_dtype(cfg.dtype, device=self._device) if self._dtype == torch.float16: raise ValueError( "full fp16 training is not supported with this recipe. Please use bf16 or fp32 instead." ) self.world_size, self.rank = utils.get_world_size_and_rank() self._is_rank_zero = self.rank == 0 # logging attributes self._output_dir = cfg.output_dir self._log_every_n_steps = cfg.get("log_every_n_steps", 1) self._log_peak_memory_stats = cfg.get("log_peak_memory_stats", False) if self._log_peak_memory_stats and self._device.type != "cuda": log.info( "log_peak_memory_stats was set to True, however, training does not use cuda. Setting log_peak_memory_stats=False." ) self._log_peak_memory_stats = False # activation checkpointing/offloading self._enable_activation_checkpointing = cfg.get( "enable_activation_checkpointing", False ) self._enable_activation_offloading = cfg.get( "enable_activation_offloading", False ) if self._enable_activation_offloading: if self._device.type != "cuda": raise RuntimeError( "enable_activation_offloading should only be True when training on CUDA" ) if not self._enable_activation_checkpointing: raise RuntimeError( "enable_activation_offloading should only be True when enable_activation_checkpointing is True" ) elif self._enable_activation_checkpointing: utils.log_rank_zero( log, "Hint: enable_activation_checkpointing is True, but enable_activation_offloading isn't. " "Enabling activation offloading should reduce memory further.", ) # These attributes constitute the recipe state and are updated by ``load_checkpoint`` # when ``resume_from_checkpoint`` is ``True`` self.seed = training.set_seed( seed=cfg.seed, debug_mode=cfg.get("cudnn_deterministic_mode", None) ) self.epochs_run = 0 self.total_epochs = cfg.epochs self.max_steps_per_epoch = cfg.max_steps_per_epoch self.global_step = 0 self._resume_from_checkpoint = cfg.resume_from_checkpoint self._save_adapter_weights_only = cfg.get("save_adapter_weights_only", False) self._gradient_accumulation_steps = cfg.gradient_accumulation_steps def load_checkpoint(self, cfg_checkpointer: DictConfig) -> Dict[str, Any]: """ Extract the checkpoint state from file and validate. This includes the base model weights. If resume_from_checkpoint is True, this also includes the adapter weights and recipe state """ self._checkpointer = config.instantiate( cfg_checkpointer, should_load_recipe_state=self._resume_from_checkpoint, ) checkpoint_dict = self._checkpointer.load_checkpoint() # When resuming from checkpoint for LoRA, the recipe expects the adapter weights # and recipe state to be present. The keys should match up with what ``save_checkpoint`` # used to create these intermediate checkpoints if self._resume_from_checkpoint: if training.ADAPTER_KEY not in checkpoint_dict: raise ValueError( "Adapter weights not found. Please ensure a valid adapter checkpoint is provided." ) # _update_recipe_state will throw an exception if the recipe state is not corrctly loaded # no need to check here self._update_recipe_state(checkpoint_dict) return checkpoint_dict def _update_recipe_state(self, ckpt_dict: Dict[str, Any]) -> None: """ Updates the recipe state from checkpoint. """ try: self.epochs_run = ckpt_dict[training.EPOCHS_KEY] # on mismatch, warn the user and prevent the override if self.seed != ckpt_dict[training.SEED_KEY]: warn( message=( "Config value for seed does not match the checkpoint value, " f"using the checkpoint value: {ckpt_dict[training.SEED_KEY]}" ) ) self.seed = ckpt_dict[training.SEED_KEY] if self.max_steps_per_epoch != ckpt_dict[training.MAX_STEPS_KEY]: warn( message=( "Config value for max_steps_per_epoch does not match the checkpoint value, " f"using the checkpoint value: {ckpt_dict[training.MAX_STEPS_KEY]}" ) ) self.max_steps_per_epoch = ckpt_dict[training.MAX_STEPS_KEY] # on mismatch, warn the user but allow the override if self.total_epochs != ckpt_dict[training.TOTAL_EPOCHS_KEY]: warn( message=( "Config value for total_epochs does not match the checkpoint value, " f"using the config value: {self.total_epochs}" ) ) except KeyError as e: raise KeyError( "Checkpoint does not contain the required keys needed for updating recipe state. " "Are you sure you passed in the right recipe checkpoint?" ) from e def setup(self, cfg: DictConfig) -> None: """ Setup the recipe state. This includes recipe state (if resume_from_checkpoint is True), model, tokenizer, loss, optimizer, learning rate scheduler, sampler, and dataloader. """ if self._is_rank_zero: self._metric_logger = config.instantiate(cfg.metric_logger) # log config with parameter override self._metric_logger.log_config(cfg) utils.log_rank_zero(log, "metric logger is initialized.") checkpoint_dict = self.load_checkpoint(cfg_checkpointer=cfg.checkpointer) self._model = self._setup_model( cfg_model=cfg.model, enable_activation_checkpointing=cfg.enable_activation_checkpointing, enable_activation_offloading=self._enable_activation_offloading, custom_sharded_layers=cfg.get("custom_sharded_layers", None), fsdp_cpu_offload=cfg.get("fsdp_cpu_offload", False), reshard_after_forward=cfg.get("fsdp_reshard_after_forward", True), base_model_state_dict=checkpoint_dict[training.MODEL_KEY], lora_weights_state_dict=( checkpoint_dict[training.ADAPTER_KEY] if self._resume_from_checkpoint else None ), ) self._tokenizer = config.instantiate(cfg.tokenizer) self._optimizer = self._setup_optimizer( cfg_optimizer=cfg.optimizer, opt_state_dict=( checkpoint_dict[training.OPT_KEY] if self._resume_from_checkpoint else None ), ) self._loss_fn = config.instantiate(cfg.loss) utils.log_rank_zero(log, "Loss is initialized.") # sampler and dataloader depend on the tokenizer and loss_fn and should be # setup after all of these are setup self._sampler, self._dataloader = self._setup_data( cfg_dataset=cfg.dataset, shuffle=cfg.shuffle, batch_size=cfg.batch_size, ) # Finally update the recipe state which can only be correctly set after all of the # other components have been initialized and updated. # Number of training steps in each epoch depends on the number of batches produced # by the dataloader and the max_steps_per_epoch param set by the user and is used # for logging and tracking training state. This should be computed after the dataloader # has been setup self._steps_per_epoch = ( len(self._dataloader) // self._gradient_accumulation_steps ) if ( self.max_steps_per_epoch is not None and self.max_steps_per_epoch < self._steps_per_epoch ): self._steps_per_epoch = self.max_steps_per_epoch self.global_step = self.epochs_run * self._steps_per_epoch # Learning rate scheduler can only be set up after number of steps # has been computed self._lr_scheduler = self._setup_lr_scheduler( cfg_lr_scheduler=cfg.lr_scheduler, num_training_steps=self.total_epochs * self._steps_per_epoch, last_epoch=self.global_step - 1, ) def _setup_model( self, cfg_model: DictConfig, enable_activation_checkpointing: bool, enable_activation_offloading: bool, fsdp_cpu_offload: bool, reshard_after_forward: bool, base_model_state_dict: Dict[str, Any], custom_sharded_layers: Optional[List[str]] = None, lora_weights_state_dict: Optional[Dict[str, Any]] = None, ) -> nn.Module: """ Model initialization has some important considerations: a. To minimize GPU peak memory, we initialize the model on meta device with the right dtype b. All ranks calls ``load_state_dict`` without peaking CPU RAMs since full state dicts are loaded with ``torch.load(mmap=True)`` c. We register (pre-)forward hooks with ``fully_shard`` instead of wrapping `nn.Module` """ self._lora_rank = cfg_model.lora_rank self._lora_alpha = cfg_model.lora_alpha self._lora_attn_modules = list(cfg_model.lora_attn_modules) self._apply_lora_to_mlp = cfg_model.apply_lora_to_mlp self._apply_lora_to_output = getattr(cfg_model, "apply_lora_to_output", False) init_start = time.perf_counter() utils.log_rank_zero( log, "FSDP is enabled. Instantiating model and loading checkpoint on Rank 0 ...", ) with training.set_default_dtype(self._dtype), torch.device("meta"): model = config.instantiate(cfg_model) self.adapter_params = get_adapter_params(model) set_trainable_params(model, self.adapter_params) if enable_activation_checkpointing: training.set_activation_checkpointing( model, auto_wrap_policy={modules.TransformerSelfAttentionLayer} ) # For FSDP sharding fsdp_shard_conditions = [ partial( training.get_shard_conditions, names_to_match=custom_sharded_layers, ) ] training.shard_model( model=model, shard_conditions=fsdp_shard_conditions, cpu_offload=fsdp_cpu_offload, reshard_after_forward=reshard_after_forward, ) if lora_weights_state_dict: lora_missing, lora_unexpected = training.load_from_full_model_state_dict( model, lora_weights_state_dict, self._device, cpu_offload=fsdp_cpu_offload, ) else: lora_missing, lora_unexpected = None, None # Initialize LoRA params and RoPE buffers with training.set_default_dtype(self._dtype), self._device: lora_device = "cpu" if fsdp_cpu_offload else self._device for m in model.modules(): if ( isinstance(m, LoRALinear) or isinstance(m, DoRALinear) ) and not lora_weights_state_dict: # lora may not be covered in state dict # if finetune for the 1st time m.to_empty(device=lora_device) m.initialize_parameters() # RoPE is not covered in state dict if hasattr(m, "rope_init"): m.rope_init() base_missing, base_unexpected = training.load_from_full_model_state_dict( model, base_model_state_dict, self._device, cpu_offload=fsdp_cpu_offload, ) is_dora = False for m in model.modules(): if hasattr(m, "initialize_dora_magnitude"): is_dora = True m.initialize_dora_magnitude() if is_dora: for m in model.modules(): if hasattr(m, "initialize_dora_magnitude"): m.initialize_dora_magnitude() validate_missing_and_unexpected_for_lora( lora_attn_modules=self._lora_attn_modules, apply_lora_to_mlp=self._apply_lora_to_mlp, apply_lora_to_output=self._apply_lora_to_output, base_missing=base_missing, base_unexpected=base_unexpected, lora_missing=lora_missing, lora_unexpected=lora_unexpected, ) # Ensure no params and buffers are on meta device # activation offloading self.activations_handling_ctx = training.get_act_offloading_ctx_manager( model, enable_activation_offloading ) training.validate_no_params_on_meta_device(model) utils.log_rank_zero( log, f"Instantiating model and loading checkpoint took {time.perf_counter() - init_start:.2f} secs", ) if self._is_rank_zero: memory_stats = training.get_memory_stats(device=self._device) training.log_memory_stats(memory_stats) # synchronize before training begins torch.distributed.barrier() return model def _setup_optimizer( self, cfg_optimizer: DictConfig, opt_state_dict: Optional[Dict[str, Any]] = None ) -> Optimizer: optimizer = config.instantiate(cfg_optimizer, self._model.parameters()) if opt_state_dict: training.load_from_full_optimizer_state_dict( self._model, optimizer, opt_state_dict, self._device, ) utils.log_rank_zero(log, "Optimizer and loss are initialized.") return optimizer def _setup_lr_scheduler( self, cfg_lr_scheduler: DictConfig, num_training_steps: int, last_epoch: int, ) -> Optimizer: lr_scheduler = config.instantiate( cfg_lr_scheduler, self._optimizer, num_training_steps=num_training_steps, last_epoch=last_epoch, ) utils.log_rank_zero(log, "Learning rate scheduler is initialized.") return lr_scheduler def _setup_data( self, cfg_dataset: DictConfig, shuffle: bool, batch_size: int, ) -> Tuple[DistributedSampler, DataLoader]: """ All data related setup happens here. Currently this recipe only supports the DistributedSamplers with Map-style Datasets which fit into memory. Other samplers, iterable datasets and streaming datasets are not supported. """ if isinstance(cfg_dataset, ListConfig): datasets = [ config.instantiate(single_cfg_dataset, tokenizer=self._tokenizer) for single_cfg_dataset in cfg_dataset ] ds = ConcatDataset(datasets=datasets) else: ds = config.instantiate(cfg_dataset, tokenizer=self._tokenizer) sampler = DistributedSampler( ds, num_replicas=self.world_size, rank=self.rank, shuffle=shuffle, seed=0 ) dataloader = DataLoader( dataset=ds, batch_size=batch_size, sampler=sampler, # dropping last avoids shape issues with compile + flex attention drop_last=True, collate_fn=partial( padded_collate_dpo, padding_idx=self._tokenizer.pad_id, ignore_idx=CROSS_ENTROPY_IGNORE_IDX, ), ) utils.log_rank_zero(log, "Dataset and Sampler are initialized.") return sampler, dataloader def save_checkpoint( self, epoch: int, ) -> None: """ Checkpoint the state of the recipe. The constructed checkpoint state dict contains the following information: - Merged weights with key MODEL_KEY - Adapter weights with key ADAPTER_KEY - Relevant recipe state if training is not complete - If the `self._save_adapter_weights_only` option is True, the checkpointer will save only the adapter weights Checkpointer will save the merged weights, adapter weights and recipe state in different checkpoint files. To correctly resume from training, the adapter weights and recipe state must be provided along with the base model weights.""" # final dict passed onto the checkpointer checkpoint_dict = {} intermediate_checkpoint = epoch + 1 < self.total_epochs # To prevent GPU memory from spiking during checkpoint save, # we consolidate the full model and optim state dicts on CPU for rank 0 cpu_state_dict = training.gather_cpu_state_dict( self._model, self._is_rank_zero, device=self._device, adapter_weights_only=self._save_adapter_weights_only, ) if intermediate_checkpoint: opt_state_dict = training.get_full_optimizer_state_dict( self._model, self._optimizer, self._is_rank_zero, device=self._device, ) else: opt_state_dict = None # Now that we have the model and opt state dict, create the actual checkpoint dict # to be sent to the checkpointer and ultimately written to file if self._is_rank_zero: if self._save_adapter_weights_only: adapter_state_dict = cpu_state_dict else: # Filter out the adapter keys and weights from the model state dict. These will # be saved separately adapter_state_dict = get_adapter_state_dict(cpu_state_dict) # merge the adapter weights and base weights to create the model checkpoint merged_state_dict = get_merged_lora_ckpt( cpu_state_dict, rank=self._lora_rank, alpha=self._lora_alpha, ) checkpoint_dict.update({training.MODEL_KEY: merged_state_dict}) checkpoint_dict.update({training.ADAPTER_KEY: adapter_state_dict}) # if training is in-progress, checkpoint the optimizer state and recipe state # as well. if intermediate_checkpoint: checkpoint_dict.update( { training.OPT_KEY: opt_state_dict, training.SEED_KEY: self.seed, training.EPOCHS_KEY: self.epochs_run, training.TOTAL_EPOCHS_KEY: self.total_epochs, training.MAX_STEPS_KEY: self.max_steps_per_epoch, } ) adapter_config = { "r": self._lora_rank, "lora_alpha": self._lora_alpha, "target_modules": get_lora_module_names( self._lora_attn_modules, self._apply_lora_to_mlp, self._apply_lora_to_output, ), "peft_type": "LORA", } checkpoint_dict.update({training.ADAPTER_CONFIG: adapter_config}) self._checkpointer.save_checkpoint( checkpoint_dict, epoch=epoch, intermediate_checkpoint=intermediate_checkpoint, adapter_only=self._save_adapter_weights_only, ) def concatenated_forward( self, model: nn.Module, batch: Tuple[torch.Tensor, torch.Tensor] ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: """ Run forward pass of the model with chosen and rejected samples concatenated. Args: model (nn.Module): The model to be used for the forward pass. batch (Tuple[torch.Tensor, torch.Tensor]): Tuple of input_ids and labels. Returns: Tuple of chosen log probs, rejected log probs, chosen logits, rejected logits. """ concatenated_input_ids, concatenated_labels = batch concatenated_input_ids = concatenated_input_ids.to(self._device) concatenated_labels = concatenated_labels.to(self._device) # formed by concatenating an equal number of "chosen" and "rejected". len_chosen = concatenated_input_ids.shape[0] // 2 with self.activations_handling_ctx: all_logits = model(concatenated_input_ids) all_log_probs = rlhf.get_batch_log_probs(all_logits, concatenated_labels) chosen_log_probs = all_log_probs[:len_chosen] rejected_log_probs = all_log_probs[len_chosen:] chosen_logits = all_logits[:len_chosen] rejected_logits = all_logits[len_chosen:] return (chosen_log_probs, rejected_log_probs, chosen_logits, rejected_logits) def train(self) -> None: """ The core training loop. """ # clean up before training begins training.cleanup_before_training() # zero out the gradients before starting training self._optimizer.zero_grad() # Initialize tokens count and running loss (for grad accumulation) t0 = time.perf_counter() # Running metrics running_loss = 0 running_metrics = { "rewards/chosen": 0, "rewards/rejected": 0, "rewards/accuracies": 0, "log_probs/chosen": 0, "log_probs/rejected": 0, "logits/chosen": 0, "logits/rejected": 0, } num_tokens = 0 # self.epochs_run should be non-zero when we're resuming from a checkpoint for curr_epoch in range(self.epochs_run, self.total_epochs): # Update the sampler to ensure data is correctly shuffled across epochs # in case shuffle is True self._sampler.set_epoch(curr_epoch) pbar = tqdm(total=self._steps_per_epoch, disable=not (self.rank == 0)) for idx, batch in enumerate(self._dataloader): if ( self.max_steps_per_epoch is not None and (idx // self._gradient_accumulation_steps) == self.max_steps_per_epoch ): break # batch is input_ids, labels num_tokens += torch.tensor(batch[0].numel()) ( policy_chosen_log_probs, policy_rejected_log_probs, policy_chosen_logits, policy_rejected_logits, ) = self.concatenated_forward(self._model, batch) policy_chosen_logits_mean = policy_chosen_logits.detach().mean() policy_rejected_logits_mean = policy_rejected_logits.detach().mean() # deleting logits here helps reduce (peak) memory usage - we only need them for metric logging del policy_chosen_logits, policy_rejected_logits with torch.no_grad(), disable_adapter(self._model): ( reference_chosen_log_probs, reference_rejected_log_probs, _, _, ) = self.concatenated_forward(self._model, batch) loss, chosen_rewards, rejected_rewards = self._loss_fn( policy_chosen_log_probs, policy_rejected_log_probs, reference_chosen_log_probs, reference_rejected_log_probs, ) reward_accuracies = (chosen_rewards > rejected_rewards).float() loss = loss.mean() loss = loss / self._gradient_accumulation_steps # Update running metrics running_loss += loss scaling_factor = ( 1 / self._gradient_accumulation_steps ) # to average out between grad_acc steps running_metrics["rewards/chosen"] += ( scaling_factor * chosen_rewards.mean() ) running_metrics["rewards/rejected"] += ( scaling_factor * rejected_rewards.mean() ) running_metrics["rewards/accuracies"] += ( scaling_factor * reward_accuracies.mean() ) running_metrics["log_probs/chosen"] += ( scaling_factor * policy_chosen_log_probs.detach().mean() ) running_metrics["log_probs/rejected"] += ( scaling_factor * policy_rejected_log_probs.detach().mean() ) running_metrics["logits/chosen"] += ( scaling_factor * policy_chosen_logits_mean ) running_metrics["logits/rejected"] += ( scaling_factor * policy_rejected_logits_mean ) loss.backward() # Step with optimizer if (idx + 1) % self._gradient_accumulation_steps == 0: # Accumulate running metrics across all devices torch.distributed.all_reduce(running_loss) torch.distributed.all_reduce(num_tokens) for key in running_metrics: torch.distributed.all_reduce( running_metrics[key], op=torch.distributed.ReduceOp.AVG ) self._optimizer.step() self._optimizer.zero_grad(set_to_none=True) self._lr_scheduler.step() # Update the number of steps when the weights are updated self.global_step += 1 loss_to_log = running_loss.item() pbar.update(1) pbar.set_description( f"{curr_epoch + 1}|{self.global_step}|Loss: {loss_to_log}" ) # Log per-step metrics if ( self.global_step % self._log_every_n_steps == 0 and self._is_rank_zero ): time_per_step = time.perf_counter() - t0 log_dict = { "loss": loss_to_log, "lr": self._optimizer.param_groups[0]["lr"], "tokens_per_second_per_gpu": num_tokens / (time_per_step * self.world_size), "rewards/chosen": running_metrics["rewards/chosen"].cpu(), "rewards/rejected": running_metrics[ "rewards/rejected" ].cpu(), "rewards/accuracies": running_metrics[ "rewards/accuracies" ].cpu(), "rewards/margins": ( running_metrics["rewards/chosen"] - running_metrics["rewards/rejected"] ).cpu(), "log_probs/chosen": running_metrics[ "log_probs/chosen" ].cpu(), "log_probs/rejected": running_metrics[ "log_probs/rejected" ].cpu(), "logits/chosen": running_metrics["logits/chosen"].cpu(), "logits/rejected": running_metrics["logits/rejected"].cpu(), } if self._log_peak_memory_stats: log_dict.update( training.get_memory_stats(device=self._device) ) self._metric_logger.log_dict( log_dict, step=self.global_step, ) # Reset running stats for the next step running_loss = 0 running_metrics = {key: 0 for key in running_metrics} num_tokens = 0 t0 = time.perf_counter() self.epochs_run += 1 self.save_checkpoint(epoch=curr_epoch) def cleanup(self) -> None: if self._is_rank_zero: self._metric_logger.close() destroy_process_group() @config.parse def recipe_main(cfg: DictConfig) -> None: """ Entry point for the recipe. Configurable parameters are read in the following order: - Parameters specified in config (see available configs through ``tune ls``) - Overwritten by arguments from the command-line """ if not training.is_distributed(): raise RuntimeError( "Distributed finetune recipe should be run via a distributed launcher." "If using tune CLI, please specify --nnodes 1 and --nproc_per_node [num_gpus]" ) init_process_group("cuda:nccl,cpu:gloo") if cfg.get("fsdp_cpu_offload", False): # Utilize all available CPU cores for intra-op parallelism. This provides ~2x # speed up when benchmarking fused AdamW on CPU training.set_torch_num_threads() config.log_config(recipe_name="LoRADPORecipeDistributed", cfg=cfg) recipe = LoRADPORecipeDistributed(cfg=cfg) recipe.setup(cfg=cfg) recipe.train() recipe.cleanup() if __name__ == "__main__": sys.exit(recipe_main())