# Copyright 2020-2025 The HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import copy import inspect import os import re import textwrap import warnings from collections import defaultdict, deque from contextlib import nullcontext from functools import partial from pathlib import Path from typing import Any, Callable, Optional, Union import datasets import torch import torch.utils.data import transformers from accelerate import logging from accelerate.utils import broadcast_object_list, gather, gather_object, is_peft_model, set_seed from datasets import Dataset, IterableDataset from torch import nn from torch.distributed.fsdp import FullyShardedDataParallel as FSDP from torch.utils.data import DataLoader, Sampler from transformers import ( AutoConfig, AutoModelForSequenceClassification, AutoProcessor, AutoTokenizer, GenerationConfig, PreTrainedModel, PreTrainedTokenizerBase, ProcessorMixin, Trainer, TrainerCallback, is_wandb_available, ) from transformers.trainer_utils import seed_worker from transformers.utils import is_datasets_available, is_flash_attn_2_available, is_peft_available, is_rich_available from ..data_utils import apply_chat_template, is_conversational, maybe_apply_chat_template from ..extras.profiling import profiling_context, profiling_decorator from ..extras.vllm_client import VLLMClient from ..import_utils import is_vllm_available from ..models import prepare_deepspeed, prepare_fsdp, prepare_peft_model, unwrap_model_for_generation from .callbacks import SyncRefModelCallback from .rloo_config import RLOOConfig from .utils import ( RepeatSampler, disable_dropout_in_model, entropy_from_logits, generate_model_card, get_comet_experiment_url, identity, nanmax, nanmin, nanstd, pad, print_prompt_completions_sample, selective_log_softmax, shuffle_sequence_dict, split_pixel_values_by_grid, split_tensor_dict, truncate_with_protected_tokens, unsplit_pixel_values_by_grid, ) if is_peft_available(): from peft import PeftConfig, PeftModel if is_vllm_available(): from vllm import LLM, SamplingParams from vllm.sampling_params import GuidedDecodingParams if is_wandb_available(): import wandb logger = logging.get_logger(__name__) # What we call a reward function is a callable that takes a list of prompts and completions and returns a list of # rewards. When it's a string, it's a model ID, so it's loaded as a pretrained model. RewardFunc = Union[str, PreTrainedModel, Callable[[list, list], list[float]]] class RLOOTrainer(Trainer): """ Trainer for the Reinforce Leave One Out (RLOO) method. This algorithm was initially proposed in the paper [Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs] (https://huggingface.co/papers/2402.14740). Example: ```python from datasets import load_dataset from trl import RLOOTrainer dataset = load_dataset("trl-lib/tldr", split="train") def reward_func(completions, **kwargs): # Dummy reward function that rewards completions with more unique letters. return [float(len(set(completion))) for completion in completions] trainer = RLOOTrainer( model="Qwen/Qwen2-0.5B-Instruct", reward_funcs=reward_func, train_dataset=dataset, ) trainer.train() ``` Args: model (`Union[str, PreTrainedModel]`): Model to be trained. Can be either: - A string, being the *model id* of a pretrained model hosted inside a model repo on huggingface.co, or a path to a *directory* containing model weights saved using [`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded using [`~transformers.AutoModelForCausalLM.from_pretrained`] with the keyword arguments in `args.model_init_kwargs`. - A [`~transformers.PreTrainedModel`] object. Only causal language models are supported. reward_funcs (`Union[RewardFunc, list[RewardFunc]]`): Reward functions to be used for computing the rewards. To compute the rewards, we call all the reward functions with the prompts and completions and sum the rewards. Can be either: - A single reward function, such as: - A string: The *model ID* of a pretrained model hosted inside a model repo on huggingface.co, or a path to a *directory* containing model weights saved using [`~transformers.PreTrainedModel.save_pretrained`], e.g., `'./my_model_directory/'`. The model is loaded using [`~transformers.AutoModelForSequenceClassification.from_pretrained`] with `num_labels=1` and the keyword arguments in `args.model_init_kwargs`. - A [`~transformers.PreTrainedModel`] object: Only sequence classification models are supported. - A custom reward function: The function is provided with the prompts and the generated completions, plus any additional columns in the dataset. It should return a list of rewards. Custom reward functions can also return `None` when the reward is not applicable to those samples. This is useful for multi-task training where different reward functions apply to different types of samples. When a reward function returns `None` for a sample, that reward function is excluded from the reward calculation for that sample. For more details, see [Using a custom reward function](#using-a-custom-reward-function). The trainer's state is also passed to the reward function. The trainer's state is an instance of [`~transformers.TrainerState`] and can be accessed by accessing the `trainer_state` argument to the reward function's signature. - A list of reward functions, where each item can independently be any of the above types. Mixing different types within the list (e.g., a string model ID and a custom reward function) is allowed. args ([`RLOOConfig`], *optional*, defaults to `None`): Configuration for this trainer. If `None`, a default configuration is used. train_dataset ([`~datasets.Dataset`] or [`~datasets.IterableDataset`]): Dataset to use for training. It must include a column `"prompt"`. Any additional columns in the dataset is ignored. The format of the samples can be either: - [Standard](dataset_formats#standard): Each sample contains plain text. - [Conversational](dataset_formats#conversational): Each sample contains structured messages (e.g., role and content). eval_dataset ([`~datasets.Dataset`], [`~datasets.IterableDataset`] or `dict[str, Union[Dataset, IterableDataset]]`): Dataset to use for evaluation. It must meet the same requirements as `train_dataset`. processing_class ([`~transformers.PreTrainedTokenizerBase`], [`~transformers.ProcessorMixin`] or `None`, *optional*, defaults to `None`): Processing class used to process the data. The padding side must be set to "left". If `None`, the processing class is loaded from the model's name with [`~transformers.AutoProcessor.from_pretrained`]. A padding token, `tokenizer.pad_token`, must be set. If the processing class has not set a padding token, `tokenizer.eos_token` will be used as the default. reward_processing_classes (`Union[PreTrainedTokenizerBase, list[PreTrainedTokenizerBase]]`, *optional*, defaults to `None`): Processing classes corresponding to the reward functions specified in `reward_funcs`. Can be either: - A single processing class: Used when `reward_funcs` contains only one reward function. - A list of processing classes: Must match the order and length of the reward functions in `reward_funcs`. If set to `None`, or if an element of the list corresponding to a [`~transformers.PreTrainedModel`] is `None`, the tokenizer for the model is automatically loaded using [`~transformers.AutoTokenizer.from_pretrained`]. For elements in `reward_funcs` that are custom reward functions (not [`~transformers.PreTrainedModel`]), the corresponding entries in `reward_processing_classes` are ignored. callbacks (list of [`~transformers.TrainerCallback`], *optional*, defaults to `None`): List of callbacks to customize the training loop. Will add those to the list of default callbacks detailed in [here](https://huggingface.co/docs/transformers/main_classes/callback). If you want to remove one of the default callbacks used, use the [`~transformers.Trainer.remove_callback`] method. optimizers (`tuple[torch.optim.Optimizer, torch.optim.lr_scheduler.LambdaLR]`, *optional*, defaults to `(None, None)`): A tuple containing the optimizer and the scheduler to use. Will default to an instance of [`AdamW`] on your model and a scheduler given by [`get_linear_schedule_with_warmup`] controlled by `args`. peft_config ([`~peft.PeftConfig`], *optional*, defaults to `None`): PEFT configuration used to wrap the model. If `None`, the model is not wrapped. """ _tag_names = ["trl", "rloo"] def __init__( self, # Note for dev: we can remove the default None when we remove the deprecated model parameter in version 0.25.0 model: Union[str, PreTrainedModel] = None, reward_funcs: Union[RewardFunc, list[RewardFunc]] = None, args: Optional[RLOOConfig] = None, train_dataset: Optional[Union[Dataset, IterableDataset]] = None, eval_dataset: Optional[Union[Dataset, IterableDataset, dict[str, Union[Dataset, IterableDataset]]]] = None, processing_class: Optional[Union[PreTrainedTokenizerBase, ProcessorMixin]] = None, reward_processing_classes: Optional[Union[PreTrainedTokenizerBase, list[PreTrainedTokenizerBase]]] = None, callbacks: Optional[list[TrainerCallback]] = None, optimizers: tuple[Optional[torch.optim.Optimizer], Optional[torch.optim.lr_scheduler.LambdaLR]] = (None, None), peft_config: Optional["PeftConfig"] = None, # Deprecated parameters config=None, reward_model=None, policy=None, ref_policy=None, data_collator=None, ): # Handle deprecated parameters if config is not None: warnings.warn( "Parameter 'config' is deprecated and will be removed in version 0.25.0. Please use 'args' instead. " "We are setting args=config" ) if args is None: args = config else: raise ValueError("Cannot specify both 'config' (deprecated) and 'args'. Please use 'args' only.") if reward_model is not None: warnings.warn( "Parameter 'reward_model' is deprecated and will be removed in version 0.25.0. Please use " "'reward_funcs' instead. We are setting reward_funcs=reward_model" ) if reward_funcs is None: reward_funcs = reward_model else: raise ValueError( "Cannot specify both 'reward_model' (deprecated) and 'reward_funcs'. Please use 'reward_funcs' " "only." ) if policy is not None: warnings.warn( "Parameter 'policy' is deprecated and will be removed in version 0.25.0. Please use 'model' instead. " "We are setting model=policy" ) if model is None: model = policy else: raise ValueError("Cannot specify both 'policy' (deprecated) and 'model'. Please use 'model' only.") if ref_policy is not None: warnings.warn( "Parameter 'ref_policy' is deprecated and will be removed in version 0.25.0. To use the initial model " "as the reference model, simply omit this parameter. The parameter is ignored." ) if data_collator is not None: warnings.warn( "Parameter 'data_collator' is deprecated and will be removed in version 0.25.0. The RLOOTrainer does " "not use a data collator, so this parameter is ignored." ) if "input_ids" in train_dataset.column_names: warnings.warn( "The training dataset contains a column named 'input_ids', indicating that it is pre-tokenized. " "Support for pre-tokenized datasets is deprecated and will be removed in version 0.25. Please provide " "the raw dataset (conversational or standard) with a 'prompt' column instead." ) def decode(example, tokenizer): return {"prompt": tokenizer.decode(example["input_ids"])} train_dataset = train_dataset.map(decode, fn_kwargs={"tokenizer": processing_class}) if eval_dataset is not None and "input_ids" in eval_dataset.column_names: warnings.warn( "The evaluation dataset contains a column named 'input_ids', indicating that it is pre-tokenized. " "Support for pre-tokenized datasets is deprecated and will be removed in version 0.25. Please provide " "the raw dataset (conversational or standard) with a 'prompt' column instead." ) def decode(example, tokenizer): return {"prompt": tokenizer.decode(example["input_ids"])} eval_dataset = eval_dataset.map(decode, fn_kwargs={"tokenizer": processing_class}) # Args if args is None: model_name = model if isinstance(model, str) else model.config._name_or_path model_name = model_name.split("/")[-1] args = RLOOConfig(f"{model_name}-RLOO") # Models # Trained model model_init_kwargs = args.model_init_kwargs or {} if isinstance(model, str): model_id = model dtype = model_init_kwargs.get("dtype") if isinstance(dtype, torch.dtype) or dtype == "auto" or dtype is None: pass # dtype is already a torch.dtype or "auto" or None elif isinstance(dtype, str): # it's a str, but not "auto" dtype = getattr(torch, dtype) model_init_kwargs["dtype"] = dtype else: raise ValueError( "Invalid `dtype` passed to `RLOOConfig`. Expected either 'auto' or a string representing " f"a `torch.dtype` (e.g., 'float32'), but got {dtype}." ) # Disable caching if gradient checkpointing is enabled (not supported) config = AutoConfig.from_pretrained(model_id) architecture = getattr(transformers, config.architectures[0]) model = architecture.from_pretrained(model_id, **model_init_kwargs) else: model_id = model.config._name_or_path if args.model_init_kwargs is not None: logger.warning( "You passed `model_init_kwargs` to the `RLOOConfig`, but your model is already instantiated. " "The `model_init_kwargs` will be ignored." ) # Some models (SmolVLM/Idefics3) don't support `logits_to_keep` argument and error out if we pass it # Inspect the forward method before we wrap the model with PEFT self.model_kwarg_keys = ( inspect.signature(model.forward).parameters.keys() if not hasattr(model, "get_base_model") else inspect.signature(model.get_base_model().forward).parameters.keys() ) if peft_config is not None or (is_peft_available() and isinstance(model, PeftModel)): model = prepare_peft_model(model, peft_config, args) # Processing class if processing_class is None: processing_class = AutoProcessor.from_pretrained(model.config._name_or_path) # Handle pad token for processors or tokenizers if isinstance(processing_class, ProcessorMixin): tokenizer = processing_class.tokenizer elif isinstance(processing_class, PreTrainedTokenizerBase): tokenizer = processing_class else: raise TypeError("The `processing_class` must be either a `PreTrainedTokenizerBase` or a `ProcessorMixin`") if tokenizer.pad_token is None: tokenizer.pad_token = tokenizer.eos_token self.pad_token = tokenizer.pad_token self.pad_token_id = tokenizer.pad_token_id self.eos_token_id = tokenizer.eos_token_id # Reward functions if not isinstance(reward_funcs, list): reward_funcs = [reward_funcs] self.reward_func_names = [] for i, reward_func in enumerate(reward_funcs): if isinstance(reward_func, str): reward_funcs[i] = AutoModelForSequenceClassification.from_pretrained( reward_func, num_labels=1, **model_init_kwargs ) if isinstance(reward_funcs[i], nn.Module): # Use Module over PretrainedModel for compat w/ compiled models self.reward_func_names.append(reward_funcs[i].config._name_or_path.split("/")[-1]) else: self.reward_func_names.append(reward_funcs[i].__name__) self.reward_funcs = reward_funcs # Reward weights if args.reward_weights is not None: if len(args.reward_weights) != len(reward_funcs): raise ValueError( f"Number of reward weights ({len(args.reward_weights)}) must match number of reward " f"functions ({len(reward_funcs)})" ) self.reward_weights = torch.tensor(args.reward_weights, dtype=torch.float32) else: self.reward_weights = torch.ones(len(reward_funcs), dtype=torch.float32) # Reward processing class if reward_processing_classes is None: reward_processing_classes = [None] * len(reward_funcs) elif not isinstance(reward_processing_classes, list): reward_processing_classes = [reward_processing_classes] if len(reward_processing_classes) != len(reward_funcs): raise ValueError( f"The number of reward processing classes ({len(reward_processing_classes)}) must match the number of " f"reward functions ({len(reward_funcs)})." ) for i, (reward_processing_class, reward_func) in enumerate(zip(reward_processing_classes, reward_funcs)): if isinstance(reward_func, PreTrainedModel): if reward_processing_class is None: reward_processing_class = AutoTokenizer.from_pretrained(reward_func.config._name_or_path) if reward_processing_class.pad_token_id is None: reward_processing_class.pad_token = reward_processing_class.eos_token # The reward model computes the reward for the latest non-padded token in the input sequence. # So it's important to set the pad token ID to the padding token ID of the processing class. reward_func.config.pad_token_id = reward_processing_class.pad_token_id reward_processing_classes[i] = reward_processing_class self.reward_processing_classes = reward_processing_classes # Training arguments self.max_prompt_length = args.max_prompt_length self.max_completion_length = args.max_completion_length self.num_generations = args.num_generations self.temperature = args.temperature self.top_p = args.top_p self.top_k = args.top_k self.min_p = args.min_p self.repetition_penalty = args.repetition_penalty self.use_transformers_paged = args.use_transformers_paged self.use_vllm = args.use_vllm self.vllm_mode = args.vllm_mode self.vllm_gpu_memory_utilization = args.vllm_gpu_memory_utilization # only applies to colocation mode self.vllm_tensor_parallel_size = args.vllm_tensor_parallel_size # only applies to colocation mode self.normalize_advantages = args.normalize_advantages self.mask_truncated_completions = args.mask_truncated_completions self.reward_clip_range = args.reward_clip_range # Datasets self.shuffle_dataset = args.shuffle_dataset if ( isinstance(train_dataset, IterableDataset) or isinstance(eval_dataset, IterableDataset) or ( isinstance(eval_dataset, dict) and any(isinstance(ds, IterableDataset) for ds in eval_dataset.values()) ) ): # See https://github.com/huggingface/trl/issues/3213 raise NotImplementedError( "Iterable datasets are not yet supported in RLOOTrainer. Please use a standard dataset instead." ) # Multi-step self.num_iterations = args.num_iterations self.epsilon_low = args.epsilon self.epsilon_high = args.epsilon_high if args.epsilon_high is not None else args.epsilon # Tracks the number of iterations (forward + backward passes), including those within a grad accum cycle self._step = 0 # Buffer the batch to reuse generated outputs across multiple updates. For more details, see # `_get_train_sampler` and `_prepare_inputs`. self._buffered_inputs = None # The trainer estimates the number of FLOPs (floating-point operations) using the number of elements in the # input tensor associated with the key "input_ids". However, in RLOO, the sampled data does not include the # "input_ids" key. Instead, the available keys is "prompt". As a result, the trainer issues the warning: # "Could not estimate the number of tokens of the input, floating-point operations will not be computed." To # suppress this warning, we set the "estimate_tokens" key in the model's "warnings_issued" dictionary to True. # This acts as a flag to indicate that the warning has already been issued. model.warnings_issued["estimate_tokens"] = True super().__init__( model=model, args=args, data_collator=identity, # No data collation is needed in RLOO train_dataset=train_dataset, eval_dataset=eval_dataset, processing_class=processing_class, callbacks=callbacks, optimizers=optimizers, ) # Reference model self.beta = args.beta if self.beta == 0.0: # If beta is 0.0, the reference model is not needed self.ref_model = None elif is_peft_model(model): # If PEFT is used, the reference model is not needed since the adapter can be disabled # to revert to the initial model. self.ref_model = None else: # For deepspeed, fsdp or non-distributed models, create a reference model from scratch config = AutoConfig.from_pretrained(model_id) architecture = getattr(transformers, config.architectures[0]) self.ref_model = architecture.from_pretrained(model_id, **model_init_kwargs) # Disable dropout in the models if args.disable_dropout: disable_dropout_in_model(model) if self.ref_model is not None: disable_dropout_in_model(self.ref_model) # Initialize the metrics self._metrics = {"train": defaultdict(list), "eval": defaultdict(list)} self._total_train_tokens = 0 self.log_completions = args.log_completions self.wandb_log_unique_prompts = args.wandb_log_unique_prompts self.num_completions_to_print = args.num_completions_to_print # Keep logs sized to the generation batch to record only outputs from the latest model update. self._logs = { "prompt": deque(maxlen=args.generation_batch_size), "completion": deque(maxlen=args.generation_batch_size), "rewards": defaultdict(lambda: deque(maxlen=args.generation_batch_size)), "advantages": deque(maxlen=args.generation_batch_size), } # Ensure each process receives a unique seed to prevent duplicate completions when generating with # transformers if num_generations exceeds per_device_train_batch_size. We could skip it if we use vLLM, but # it's safer to set it in all cases. set_seed(args.seed, device_specific=True) if self.use_vllm: if not is_vllm_available(): raise ImportError( "vLLM is not available and `use_vllm` is set to True. Please install vLLM with " "`pip install vllm` to use it." ) if self.vllm_mode == "server": if self.accelerator.is_main_process: if args.vllm_server_base_url is not None: base_url = args.vllm_server_base_url else: base_url = f"http://{args.vllm_server_host}:{args.vllm_server_port}" self.vllm_client = VLLMClient(base_url=base_url, connection_timeout=args.vllm_server_timeout) self.vllm_client.init_communicator(device=torch.cuda.current_device()) elif self.vllm_mode == "colocate": # Make sure vllm_tensor_parallel_size group size evenly divides the world size - each group should have # the same number of ranks if not self.accelerator.num_processes % self.vllm_tensor_parallel_size == 0: raise ValueError( f"vllm_tensor_parallel_size ({self.vllm_tensor_parallel_size}) must divide world size " f"({self.accelerator.num_processes}) evenly." ) if self.vllm_tensor_parallel_size > 1: # Create subgroups of ranks for TP, each group with `vllm_tensor_parallel_size` ranks. # For example, if world_size=8 and vllm_tensor_parallel_size=2 → groups: [0,1], [2,3], [4,5], [6,7] self.tp_group, _ = torch.distributed.new_subgroups_by_enumeration( [ list(range(i * self.vllm_tensor_parallel_size, (i + 1) * self.vllm_tensor_parallel_size)) for i in range(self.accelerator.num_processes // self.vllm_tensor_parallel_size) ] ) # vLLM requires the environment variables to be set for distributed training. os.environ["RANK"] = str(self.accelerator.process_index) os.environ["LOCAL_RANK"] = str(self.accelerator.local_process_index) os.environ["WORLD_SIZE"] = str(self.accelerator.num_processes) os.environ["MASTER_ADDR"] = os.environ.get("MASTER_ADDR", "localhost") os.environ["MASTER_PORT"] = os.environ.get("MASTER_PORT", "12345") if self.max_prompt_length is not None and self.max_completion_length is not None: max_model_len = self.max_prompt_length + self.max_completion_length else: max_model_len = None self.llm = LLM( model=model.name_or_path, tensor_parallel_size=args.vllm_tensor_parallel_size, gpu_memory_utilization=self.vllm_gpu_memory_utilization, max_num_seqs=self.args.per_device_train_batch_size * self.vllm_tensor_parallel_size * self.args.steps_per_generation, max_model_len=max_model_len, distributed_executor_backend="external_launcher", # Feed identical seed for tp groups to ensure sampling results are the same across workers seed=self.accelerator.process_index // self.vllm_tensor_parallel_size, # Latest vLLM v1 memory profiler is misled by the high default value (i.e., 32768) - thinking there's not enough memory max_num_batched_tokens=4096, model_impl=self.args.vllm_model_impl, ) else: raise ValueError(f"vllm_mode must be either 'server' or 'colocate', got '{self.vllm_mode}'.") # vLLM specific sampling arguments self.guided_decoding_regex = args.vllm_guided_decoding_regex self._last_loaded_step = -1 # tag to avoid useless loading during grad accumulation # When using vLLM, the main process is responsible for loading the model weights. This can cause process # desynchronization and seems to lead to DeepSpeed hanging during initialization. To prevent this, we # synchronize all processes after vLLM has been fully initialized. self.accelerator.wait_for_everyone() else: generation_kwargs = { "max_new_tokens": self.max_completion_length, "do_sample": True, "pad_token_id": tokenizer.pad_token_id, "bos_token_id": tokenizer.bos_token_id, "eos_token_id": tokenizer.eos_token_id, "temperature": self.temperature, "top_p": self.top_p, "top_k": self.top_k, "min_p": self.min_p, "repetition_penalty": self.repetition_penalty, "cache_implementation": args.cache_implementation, } if args.use_transformers_paged: generation_kwargs["max_batch_tokens"] = 512 generation_kwargs["num_blocks"] = 1024 generation_kwargs["block_size"] = 128 if args.generation_kwargs is not None: generation_kwargs.update(args.generation_kwargs) self.generation_config = GenerationConfig(**generation_kwargs) # Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the # model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set # self.model_accepts_loss_kwargs to False to enable scaling. self.model_accepts_loss_kwargs = False # Add tags to the model self.model.add_model_tags(self._tag_names) if self.ref_model is not None: if self.is_deepspeed_enabled: self.ref_model = prepare_deepspeed(self.ref_model, self.accelerator) elif self.is_fsdp_enabled: self.ref_model = prepare_fsdp(self.ref_model, self.accelerator) else: self.ref_model = self.accelerator.prepare_model(self.ref_model, evaluation_mode=True) if args.sync_ref_model: self.add_callback(SyncRefModelCallback(ref_model=self.ref_model, accelerator=self.accelerator)) for i, reward_func in enumerate(self.reward_funcs): if isinstance(reward_func, PreTrainedModel): if self.is_deepspeed_enabled: self.reward_funcs[i] = prepare_deepspeed(reward_func, self.accelerator) else: # set device placement to True to make `prepare_model` move `reward_func` to device when using fsdp self.reward_funcs[i] = self.accelerator.prepare_model( reward_func, evaluation_mode=True, device_placement=True ) def _set_signature_columns_if_needed(self): # If `self.args.remove_unused_columns` is True, non-signature columns are removed. # By default, this method sets `self._signature_columns` to the model's expected inputs. # In RLOOTrainer, we preprocess data, so using the model's signature columns doesn't work. # Instead, we set them to the columns expected by the `training_step` method, hence the override. if self._signature_columns is None: self._signature_columns = ["prompt"] # This method overrides `Trainer.get_train_dataloader` to support our custom batching strategy. # Instead of returning a standard per-step batch (i.e., `per_device_batch_size), our dataloader loads an # *generation* batch (i.e., `per_device_batch_size × steps_per_generation`). This allows us to generate completions # once every steps_per_generation step—rather than once per accumulation step—which is significantly more # efficient. The only change from the original implementation is multiplying the batch size by # `steps_per_generation`. Thus, `_prepare_inputs` is called with this *generation* batch, and it handles the # splitting internally. # Maintenance note: This method is a copy-paste of the original `Trainer.get_train_dataloader` with only one line # modification. As a result, some parts of the method aren't relevant to RLOO, but we keep them to stay one line # apart from the super method, ensuring easier maintenance in the future. def get_train_dataloader(self): if self.train_dataset is None: raise ValueError("Trainer: training requires a train_dataset.") train_dataset = self.train_dataset data_collator = self.data_collator if is_datasets_available() and isinstance(train_dataset, datasets.Dataset): train_dataset = self._remove_unused_columns(train_dataset, description="training") else: data_collator = self._get_collator_with_removed_columns(data_collator, description="training") dataloader_params = { "batch_size": self._train_batch_size * self.args.steps_per_generation, # < this is the change "collate_fn": data_collator, "num_workers": self.args.dataloader_num_workers, "pin_memory": self.args.dataloader_pin_memory, "persistent_workers": self.args.dataloader_persistent_workers, } if not isinstance(train_dataset, torch.utils.data.IterableDataset): dataloader_params["sampler"] = self._get_train_sampler() dataloader_params["drop_last"] = self.args.dataloader_drop_last dataloader_params["worker_init_fn"] = partial( seed_worker, num_workers=self.args.dataloader_num_workers, rank=self.args.process_index ) dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor return self.accelerator.prepare(DataLoader(train_dataset, **dataloader_params)) def _get_train_sampler(self, dataset: Optional[Dataset] = None) -> Sampler: # Returns a sampler that # 1. ensures each prompt is repeated across multiple processes. This guarantees that identical prompts are # distributed to different GPUs, allowing rewards to be computed and normalized correctly within each prompt # group. Using the same seed across processes ensures consistent prompt assignment, preventing discrepancies # in group formation. # 2. repeats the batch multiple times to allow reusing generations across multiple updates. Refer to # _prepare_inputs to see how the generations are stored and reused. # In the following figure, the values are the prompt indices. The first row shows the first sampled batch, the # second row shows the second sampled batch, and so on. # # | GPU 0 | GPU 1 | # # global_step step <-───> num_generations=2 # <-───────> per_device_train_batch_size=3 # grad_accum ▲ ▲ 0 0 0 0 1 1 2 2 <- Generate for the first `steps_per_generation` (prompts 0 to 11); store the completions; use the first slice to compute the loss # =2 ▼ | 0 1 3 3 4 4 5 5 <- Take the stored generations and use the second slice to compute the loss # | # | 1 2 6 6 7 7 8 8 <- Take the stored generations and use the third slice to compute the loss # steps_per_gen=4 ▼ 1 3 9 9 10 10 11 11 <- Take the stored generations and use the fourth slice to compute the loss # # 2 4 12 12 13 13 14 14 <- Generate for the second `steps_per_generation` (prompts 12 to 23); store the completions; use the first slice to compute the loss # 2 5 15 15 16 16 17 17 <- Take the stored generations and use the second slice to compute the loss # ... if dataset is None: dataset = self.train_dataset return RepeatSampler( data_source=dataset, mini_repeat_count=self.num_generations, batch_size=self.args.generation_batch_size // self.num_generations, repeat_count=self.num_iterations * self.args.steps_per_generation, shuffle=self.shuffle_dataset, seed=self.args.seed, ) def _get_eval_sampler(self, eval_dataset) -> Sampler: # See _get_train_sampler for an explanation of the sampler. return RepeatSampler( data_source=eval_dataset, mini_repeat_count=self.num_generations, seed=self.args.seed, ) @profiling_decorator def _get_per_token_logps_and_entropies( self, model, input_ids, attention_mask, logits_to_keep, batch_size=None, compute_entropy=False, ) -> dict[str, Optional[torch.Tensor]]: """Compute log-probs and (optionally) entropies for each token.""" batch_size = batch_size or input_ids.size(0) # Chunk inputs into smaller batches to reduce memory peak all_logps = [] all_entropies = [] for start in range(0, input_ids.size(0), batch_size): input_ids_batch = input_ids[start : start + batch_size] attention_mask_batch = attention_mask[start : start + batch_size] # Build model inputs - check if the model supports logits_to_keep (some models and VLMs don't) model_inputs = {"input_ids": input_ids_batch, "attention_mask": attention_mask_batch} # Only add logits_to_keep if the model supports it if "logits_to_keep" in self.model_kwarg_keys: # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded model_inputs["logits_to_keep"] = logits_to_keep + 1 model_inputs["use_cache"] = False # only used in generation; set False to suppress warnings logits = model(**model_inputs).logits # Exclude the last value: it corresponds to the next token pred logits = logits[:, :-1, :] # (B, L-1, H) # Only keep the last logits_to_keep. For model that support logits_to_keep, this is a no-op. logits = logits[:, -logits_to_keep:, :] # (B, logits_to_keep, H) # Divide logits by sampling temperature. # See https://huggingface.co/blog/the_n_implementation_details_of_rlhf_with_ppo#policy-training-implementation-details logits = logits / self.temperature completion_ids = input_ids_batch[:, -logits_to_keep:] logps = selective_log_softmax(logits, completion_ids) # compute logprobs all_logps.append(logps) if compute_entropy: with torch.no_grad(): entropies = entropy_from_logits(logits) all_entropies.append(entropies) logps = torch.cat(all_logps, dim=0) entropies = torch.cat(all_entropies, dim=0) if compute_entropy else None return logps, entropies def _fix_param_name_to_vllm(self, name, extra_prefixes: Optional[list[str]] = None): extra_prefixes = extra_prefixes or [] prefixes = ["_checkpoint_wrapped_module."] + extra_prefixes for prefix in prefixes: name = name.replace(prefix, "") return name def _sync_fsdp1_params_to_vllm(self, module: nn.Module, prefix: str = "", visited=None): """Memory-efficient post-order traversal of FSDP modules to extract full parameters and sync with vLLM.""" # For FSDP1, we need to recurse into children and also use summon_full_params if visited is None: visited = set() for child_name, child_module in module.named_children(): child_prefix = f"{prefix}.{child_name}" if prefix else child_name self._sync_fsdp1_params_to_vllm( child_module, prefix=child_prefix, visited=visited ) # recurse into the child if isinstance(module, FSDP): with FSDP.summon_full_params(module, recurse=False, writeback=False): for param_name, param in module.named_parameters(): full_name = f"{prefix}.{param_name}" if prefix else param_name full_name = self._fix_param_name_to_vllm(full_name, extra_prefixes=["_fsdp_wrapped_module."]) if full_name in visited: continue # skip FSDP subtrees already traversed visited.add(full_name) if self.vllm_mode == "server" and self.accelerator.is_main_process: self.vllm_client.update_named_param(full_name, param.data) elif self.vllm_mode == "colocate": llm_model = self.llm.llm_engine.model_executor.driver_worker.model_runner.model llm_model.load_weights([(full_name, param.data)]) def _sync_fsdp2_params_to_vllm(self, module: nn.Module): # For FSDP2, module.state_dict() already covers all parameters, so no need for recursion for name, param in module.state_dict().items(): if param.is_cpu: param = param.to(torch.device("cuda")) param = param.full_tensor() if self.vllm_mode == "server" and self.accelerator.is_main_process: self.vllm_client.update_named_param(name, param) elif self.vllm_mode == "colocate": llm_model = self.llm.llm_engine.model_executor.driver_worker.model_runner.model llm_model.load_weights([(name, param)]) @profiling_decorator def _move_model_to_vllm(self): # For DeepSpeed ZeRO-3 and FSDP, we need to gather all parameters before operations deepspeed_plugin = self.accelerator.state.deepspeed_plugin zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3 if zero_stage_3: import deepspeed gather_if_zero3 = deepspeed.zero.GatheredParameters else: gather_if_zero3 = nullcontext if is_peft_model(self.model): # With PEFT and FSDP/DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as # merging adapters in a sharded manner is not supported. # TODO: does this work with FSDP? with gather_if_zero3(list(self.model.parameters())): self.model.merge_adapter() # Update vLLM weights while parameters are gathered if self.is_fsdp_enabled: # note if using FSDP, gather_if_zero3 is nullcontext # Update vLLM weights while parameters are gathered # For PEFT with FSDP we need to use the memory efficient post-order traversal fsdp_plugin = getattr(self.accelerator.state, "fsdp_plugin", None) fsdp_version = getattr(fsdp_plugin, "fsdp_version", 1) if fsdp_plugin else 1 if fsdp_version == 1: self._sync_fsdp1_params_to_vllm( self.model ) # use memory-efficient post-order traversal for FSDP elif fsdp_version == 2: self._sync_fsdp2_params_to_vllm(self.model) else: # DeepSpeed ZeRO-3 with PEFT for name, param in self.model.named_parameters(): # When using PEFT, we need to recover the original parameter name and discard some parameters name = name.removeprefix("base_model.model.").replace(".base_layer", "") if self.model.prefix in name: continue # When module to save, remove its prefix and discard the original module if "original_module" in name: continue name = self._fix_param_name_to_vllm(name, extra_prefixes=["modules_to_save.default."]) if self.vllm_mode == "server" and self.accelerator.is_main_process: self.vllm_client.update_named_param(name, param.data) elif self.vllm_mode == "colocate": llm_model = self.llm.llm_engine.model_executor.driver_worker.model_runner.model llm_model.load_weights([(name, param.data)]) # Unmerge adapters while parameters are still gathered self.model.unmerge_adapter() # Parameters will automatically be repartitioned when exiting the context else: # For non-PEFT models, simply gather (if needed) and update each parameter individually. if self.is_fsdp_enabled: fsdp_plugin = getattr(self.accelerator.state, "fsdp_plugin", None) fsdp_version = getattr(fsdp_plugin, "fsdp_version", 1) if fsdp_plugin else 1 if fsdp_version == 1: self._sync_fsdp1_params_to_vllm(self.model) # use memory-efficient post-order traversal for FSDP elif fsdp_version == 2: self._sync_fsdp2_params_to_vllm(self.model) else: for name, param in self.model.named_parameters(): name = self._fix_param_name_to_vllm(name) with gather_if_zero3([param]): if self.vllm_mode == "server" and self.accelerator.is_main_process: self.vllm_client.update_named_param(name, param.data) elif self.vllm_mode == "colocate": llm_model = self.llm.llm_engine.model_executor.driver_worker.model_runner.model llm_model.load_weights([(name, param.data)]) # Reset cache on vLLM if self.vllm_mode == "server" and self.accelerator.is_main_process: self.vllm_client.reset_prefix_cache() elif self.vllm_mode == "colocate": self.llm.reset_prefix_cache() @profiling_decorator def _prepare_inputs( self, generation_batch: dict[str, Union[torch.Tensor, Any]] ) -> dict[str, Union[torch.Tensor, Any]]: # Prepares inputs for model training/evaluation by managing completion generation and batch handling. # During training: # - Receives the local generation batch (Per-GPU batch size × steps per generation) # from the modified training dataloader instead of the standard local batch # - Generates completions once for the entire generation batch and splits it into batches of size # `per_device_train_batch_size` # - Buffers these completions and returns the appropriate slice for the current accumulation step # - Optimizes by regenerating completions only periodically (every steps_per_generation * num_iterations) # During evaluation: # - The input is treated as a standard local batch (no accumulation, no multiple iterations) # - Completions are generated for each batch without buffering or reuse # Returns a single local batch in both cases. mode = "train" if self.model.training else "eval" if mode == "train": generate_every = self.args.steps_per_generation * self.num_iterations if self._step % generate_every == 0 or self._buffered_inputs is None: # self._buffered_inputs=None can occur when resuming from a checkpoint generation_batch = self._generate_and_score_completions(generation_batch) generation_batch = split_pixel_values_by_grid(generation_batch) generation_batch = shuffle_sequence_dict(generation_batch) generation_batches = split_tensor_dict(generation_batch, self.args.steps_per_generation) self._buffered_inputs = [unsplit_pixel_values_by_grid(batch) for batch in generation_batches] inputs = self._buffered_inputs[self._step % self.args.steps_per_generation] self._step += 1 else: # In evaluation, there is neither batch grouping for generation, nor multiple iterations, hence # local generation batch == local eval batch inputs = self._generate_and_score_completions(generation_batch) return inputs @profiling_decorator def _calculate_rewards(self, inputs, prompts, completions, completion_ids_list): device = self.accelerator.device rewards_per_func = torch.zeros(len(prompts), len(self.reward_funcs), device=device) # Repeat all input columns (but "prompt", "completion", and "completion_ids") to match the num of generations keys = [key for key in inputs[0] if key not in ["prompt", "completion", "completion_ids"]] reward_kwargs = {key: [example[key] for example in inputs] for key in keys} # This allows for dynamic reward shaping based on training progress. reward_kwargs["trainer_state"] = self.state for i, (reward_func, reward_processing_class, reward_func_name) in enumerate( zip(self.reward_funcs, self.reward_processing_classes, self.reward_func_names) ): with profiling_context(self, reward_func_name): if isinstance(reward_func, nn.Module): # Module (no PretrainedModel) for compat with compiled models if is_conversational(inputs[0]): messages = [{"messages": p + c} for p, c in zip(prompts, completions)] texts = [apply_chat_template(x, reward_processing_class)["text"] for x in messages] else: texts = [p + c for p, c in zip(prompts, completions)] reward_inputs = reward_processing_class( text=texts, return_tensors="pt", padding=True, padding_side="right", add_special_tokens=False ) reward_inputs = super()._prepare_inputs(reward_inputs) with torch.inference_mode(): rewards_per_func[:, i] = reward_func(**reward_inputs).logits[:, 0] # Shape (B*G,) else: output_reward_func = reward_func( prompts=prompts, completions=completions, completion_ids=completion_ids_list, **reward_kwargs ) # Convert None values to NaN output_reward_func = [reward if reward is not None else torch.nan for reward in output_reward_func] rewards_per_func[:, i] = torch.tensor(output_reward_func, dtype=torch.float32, device=device) # If all reward functions return None for a given row, issue a detailed warning if torch.isnan(rewards_per_func).all(dim=1).any(): nan_row_idx = torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0] row_reward_kwargs = { key: value[nan_row_idx] for key, value in reward_kwargs.items() if key != "trainer_state" } row_reward_kwargs["prompt"] = prompts[nan_row_idx] row_reward_kwargs["completion"] = completions[nan_row_idx] logger.warning( f"All reward functions returned None for the following kwargs:\n{row_reward_kwargs}\n" "Please ensure that at least one reward function returns a valid reward." ) # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the # completions may be distributed across processes rewards_per_func = gather(rewards_per_func) return rewards_per_func def _generate_and_score_completions( self, inputs: list[dict[str, Union[torch.Tensor, Any]]] ) -> dict[str, Union[torch.Tensor, Any]]: device = self.accelerator.device mode = "train" if self.model.training else "eval" prompts = [x["prompt"] for x in inputs] # We don't yet support visual reward models/function, so we keep a copy of the original text-only prompts for # later use in the reward computation. If images are present, we insert {"type": "image"} as required by the # VLM chat template. original_prompts = copy.deepcopy(prompts) prompts_text = [maybe_apply_chat_template(example, self.processing_class)["prompt"] for example in inputs] prompt_inputs = self.processing_class( text=prompts_text, return_tensors="pt", padding=True, padding_side="left", add_special_tokens=False, ) prompt_inputs = super()._prepare_inputs(prompt_inputs) prompt_ids, prompt_mask = prompt_inputs["input_ids"], prompt_inputs["attention_mask"] if self.max_prompt_length is not None: # If max_prompt_length is set, we trim the prompt to keep only the last `max_prompt_length` tokens. # Then we decode those tokens back into text. We manually remove leading pad tokens from the decoded text, # because we can't use `skip_special_tokens=True` (some special tokens are still needed for generation). prompt_ids, prompt_mask = truncate_with_protected_tokens( prompt_ids, prompt_mask, self.max_prompt_length, protected_tokens=[] ) prompts_text = self.processing_class.batch_decode( prompt_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False ) prompts_text = [re.sub(rf"^({re.escape(self.pad_token)})+", "", text) for text in prompts_text] # Generate completions using either vLLM or regular generation if self.use_vllm: # First, update the vLLM weights if needed if self.state.global_step != self._last_loaded_step: self._move_model_to_vllm() self._last_loaded_step = self.state.global_step # Generate completions using vLLM: gather all prompts and use them in a single call in the main process if self.vllm_mode == "server": all_prompts_text = gather_object(prompts_text) if self.accelerator.is_main_process: # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate # num_generations outputs for each one. This is faster than generating outputs for each duplicate # prompt individually. ordered_set_of_prompts = all_prompts_text[:: self.num_generations] with profiling_context(self, "vLLM.generate"): completion_ids = self.vllm_client.generate( prompts=ordered_set_of_prompts, n=self.num_generations, repetition_penalty=self.repetition_penalty, temperature=self.temperature, top_p=self.top_p, top_k=-1 if self.top_k is None else self.top_k, min_p=0.0 if self.min_p is None else self.min_p, max_tokens=self.max_completion_length, guided_decoding_regex=self.guided_decoding_regex, generation_kwargs=self.args.generation_kwargs, ) else: completion_ids = [None] * len(all_prompts_text) # Broadcast the completions from the main process to all processes, ensuring each process receives its # corresponding slice. completion_ids = broadcast_object_list(completion_ids, from_process=0) process_slice = slice( self.accelerator.process_index * len(prompts), (self.accelerator.process_index + 1) * len(prompts), ) completion_ids = completion_ids[process_slice] # Generate completions using colocated vLLM instances: each device holds vLLM copy and work on their own batch of prompts elif self.vllm_mode == "colocate": if self.guided_decoding_regex: guided_decoding = GuidedDecodingParams(regex=self.guided_decoding_regex) else: guided_decoding = None generation_kwargs = { "n": 1, # vLLM on each GPU generates only 1 in colocate mode "repetition_penalty": self.repetition_penalty, "temperature": self.temperature, "top_p": self.top_p, "top_k": -1 if self.top_k is None else self.top_k, "min_p": 0.0 if self.min_p is None else self.min_p, "max_tokens": self.max_completion_length, "guided_decoding": guided_decoding, } if self.args.generation_kwargs is not None: generation_kwargs.update(self.args.generation_kwargs) sampling_params = SamplingParams(**generation_kwargs) if self.vllm_tensor_parallel_size > 1: # Gather prompts from all ranks in the TP group and flatten. # Each rank starts with its own prompts; after gathering, all ranks see the full group set. orig_size = len(prompts_text) gathered_prompts = [None for _ in range(self.vllm_tensor_parallel_size)] torch.distributed.all_gather_object(gathered_prompts, prompts_text, group=self.tp_group) all_prompts_text = [p for sublist in gathered_prompts for p in sublist] else: all_prompts_text = prompts_text vllm_inputs = all_prompts_text with profiling_context(self, "vLLM.generate"): all_outputs = self.llm.generate(vllm_inputs, sampling_params=sampling_params, use_tqdm=False) completion_ids = [output.token_ids for outputs in all_outputs for output in outputs.outputs] if self.vllm_tensor_parallel_size > 1: # Slice completions for this rank within its TP group. # Each rank generates all outputs — we keep only our share. local_rank_in_group = torch.distributed.get_rank(group=self.tp_group) tp_slice = slice(local_rank_in_group * orig_size, (local_rank_in_group + 1) * orig_size) completion_ids = completion_ids[tp_slice] # Pad the completions, and concatenate them with the prompts completion_ids = [torch.tensor(ids, device=device) for ids in completion_ids] completion_ids = pad(completion_ids, padding_value=self.pad_token_id) prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1) elif self.use_transformers_paged: # Re-process inputs for paged generation if needed paged_prompt_inputs = self.processing_class(text=prompts_text) previous_attn = self.model_wrapped.config._attn_implementation if is_flash_attn_2_available(): self.model_wrapped.config._attn_implementation = "paged_attention" else: self.model_wrapped.config._attn_implementation = "sdpa_paged" with ( profiling_context(self, "transformers.generate_batch"), unwrap_model_for_generation( self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation ) as unwrapped_model, torch.no_grad(), FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(), ): # Cast to the appropriate dtype based on training configuration if self.args.bf16: unwrapped_model.to(torch.bfloat16) elif self.args.fp16: unwrapped_model.to(torch.float16) with torch.inference_mode(): all_outputs = unwrapped_model.generate_batch( paged_prompt_inputs.input_ids, generation_config=self.generation_config, progress_bar=False ) completion_ids = [output.generated_tokens for output in all_outputs.values()] completion_ids = [torch.tensor(ids, device=device) for ids in completion_ids] completion_ids = pad(completion_ids, padding_value=self.pad_token_id, padding_side="right") prompt_ids = [torch.tensor(ids, device=device) for ids in paged_prompt_inputs.input_ids] prompt_ids = pad(prompt_ids, padding_value=self.pad_token_id, padding_side="left") prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1) # Restore the original attention implementation, training mode self.model_wrapped.config._attn_implementation = previous_attn else: # Regular generation path with ( profiling_context(self, "transformers.generate"), unwrap_model_for_generation( self.model_wrapped, self.accelerator, gather_deepspeed3_params=self.args.ds3_gather_for_generation ) as unwrapped_model, torch.no_grad(), FSDP.summon_full_params(self.model_wrapped, recurse=False) if self.is_fsdp_enabled else nullcontext(), ): prompt_inputs["input_ids"], prompt_inputs["attention_mask"] = prompt_ids, prompt_mask prompt_completion_ids = unwrapped_model.generate( **prompt_inputs, generation_config=self.generation_config, disable_compile=True ) # Compute prompt length and extract completion ids prompt_length = prompt_ids.size(1) prompt_ids = prompt_completion_ids[:, :prompt_length] completion_ids = prompt_completion_ids[:, prompt_length:] # Mask everything after the first EOS token is_eos = completion_ids == self.eos_token_id eos_idx = torch.full((is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device) eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)] sequence_indices = torch.arange(is_eos.size(1), device=device).expand(is_eos.size(0), -1) completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int() # Convert tensor to a list of lists of token IDs. This will be passed to the reward function, avoiding the need # to re-tokenize completions if the reward is computed from tokens. completion_ids_list = [row[mask_row].tolist() for row, mask_row in zip(completion_ids, completion_mask.bool())] # Sum along sequence dimension (dim=1) to get completion length per sequence, used for logging completion_lengths = completion_mask.sum(1) # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask if self.mask_truncated_completions: truncated_completions = ~is_eos.any(dim=1) completion_mask = completion_mask * (~truncated_completions).unsqueeze(1).int() # Concatenate prompt_mask with completion_mask for logit computation attention_mask = torch.cat([prompt_mask, completion_mask], dim=1) # (B, P+C) logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens batch_size = self.args.per_device_train_batch_size if mode == "train" else self.args.per_device_eval_batch_size with torch.no_grad(): # Compute the per-token log probabilities for the current model old_per_token_logps, _ = self._get_per_token_logps_and_entropies( self.model, prompt_completion_ids, attention_mask, logits_to_keep, batch_size, ) old_logps = (old_per_token_logps * completion_mask).sum(1) # mask out padding and tokens after EOS # Compute the per-token log probabilities for the reference model if self.beta != 0.0: if self.ref_model is not None: ref_per_token_logps, _ = self._get_per_token_logps_and_entropies( self.ref_model, prompt_completion_ids, attention_mask, logits_to_keep, batch_size=batch_size, ) else: with self.accelerator.unwrap_model(self.model).disable_adapter(): ref_per_token_logps, _ = self._get_per_token_logps_and_entropies( self.model, prompt_completion_ids, attention_mask, logits_to_keep, batch_size=batch_size, ) else: ref_per_token_logps = None # Decode the generated completions completions_text = self.processing_class.batch_decode(completion_ids, skip_special_tokens=True) if is_conversational(inputs[0]): completions = [] for prompt, completion in zip(prompts, completions_text): bootstrap = prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else "" completions.append([{"role": "assistant", "content": bootstrap + completion}]) else: completions = completions_text # Calculate rewards for each reward function. rewards_per_func aggregates rewards across all processes. This is # important because rewards will be normalized per group, and completions are distributed. We will later slice # rewards_per_func to extract each process's subset. rewards_per_func = self._calculate_rewards(inputs, original_prompts, completions, completion_ids_list) # Apply weights to each reward function's output and sum rewards = (rewards_per_func * self.reward_weights.to(device).unsqueeze(0)).nansum(dim=1) # Apply reward clipping if specified if self.reward_clip_range: rewards = rewards.clamp(min=self.reward_clip_range[0], max=self.reward_clip_range[1]) # Include the KL penalty in the reward if self.beta != 0.0: per_token_kl = old_per_token_logps - ref_per_token_logps # Apply sequence-level KL penalty to rewards (sum KL across tokens first, then apply to each sequence) kl = (per_token_kl * completion_mask).sum(-1) kl = gather(kl) # rewards are gathered, so kl must be too rewards = rewards - self.beta * kl grouped_rewards = rewards.view(-1, self.num_generations) mean_grouped_rewards = grouped_rewards.mean(dim=1) std_rewards = grouped_rewards.std(dim=1) is_std_zero = torch.isclose(std_rewards, torch.zeros_like(std_rewards)) # RLOO advantages computation grouped_sum = grouped_rewards.sum(dim=1, keepdim=True) # (num_prompts, 1) baselines = (grouped_sum - grouped_rewards) / (self.num_generations - 1) # (num_prompts, num_generations) baselines = baselines.view(-1) # Flatten back to match rewards shape advantages = rewards - baselines # Normalize advantages if self.normalize_advantages: advantages = (advantages - advantages.mean()) / (advantages.std() + 1e-4) # Slice to keep only the local part of the data process_slice = slice( self.accelerator.process_index * len(prompts), (self.accelerator.process_index + 1) * len(prompts), ) all_process_advantages = advantages.clone() # keep the aggregated advantages for logging advantages = advantages[process_slice] # Log the metrics if mode == "train": self.state.num_input_tokens_seen += self.accelerator.gather(attention_mask.sum()).sum().item() self._metrics[mode]["num_tokens"] = [self.state.num_input_tokens_seen] # Calculate and log the mean KL divergence between current and reference model if self.beta != 0.0: mean_kl = (per_token_kl * completion_mask).sum() / completion_mask.sum().clamp(min=1.0) self._metrics[mode]["kl"].append(self.accelerator.gather(mean_kl).nanmean().item()) # Log completion lengths, mean, min, max agg_completion_lengths = self.accelerator.gather(completion_lengths) self._metrics[mode]["completions/mean_length"].append(agg_completion_lengths.float().mean().item()) self._metrics[mode]["completions/min_length"].append(agg_completion_lengths.float().min().item()) self._metrics[mode]["completions/max_length"].append(agg_completion_lengths.float().max().item()) # Identify sequences that terminated with EOS and log their lengths agg_terminated_with_eos = self.accelerator.gather(is_eos.any(dim=1)) term_completion_lengths = agg_completion_lengths[agg_terminated_with_eos] clipped_completions_ratio = 1 - len(term_completion_lengths) / len(agg_completion_lengths) self._metrics[mode]["completions/clipped_ratio"].append(clipped_completions_ratio) if len(term_completion_lengths) == 0: # edge case where no terminated sequences are found term_completion_lengths = torch.zeros(1, device=device) self._metrics[mode]["completions/mean_terminated_length"].append(term_completion_lengths.float().mean().item()) self._metrics[mode]["completions/min_terminated_length"].append(term_completion_lengths.float().min().item()) self._metrics[mode]["completions/max_terminated_length"].append(term_completion_lengths.float().max().item()) # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values) for i, reward_func_name in enumerate(self.reward_func_names): mean_rewards = torch.nanmean(rewards_per_func[:, i]).item() self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards) std_func_rewards = nanstd(rewards_per_func[:, i]).item() self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_func_rewards) self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item()) self._metrics[mode]["reward_std"].append(std_rewards.mean().item()) self._metrics[mode]["frac_reward_zero_std"].append(is_std_zero.float().mean().item()) # Log prompt and completion texts self._logs["prompt"].extend(gather_object(prompts_text)) self._logs["completion"].extend(gather_object(completions_text)) for i, name in enumerate(self.reward_func_names): self._logs["rewards"][name].extend(rewards_per_func[:, i].tolist()) self._logs["advantages"].extend(all_process_advantages.tolist()) output = { "prompt_ids": prompt_ids, "prompt_mask": prompt_mask, "completion_ids": completion_ids, "completion_mask": completion_mask, "old_logps": old_logps, "advantages": advantages, } return output @profiling_decorator def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None): if return_outputs: raise ValueError("The RLOOTrainer does not support returning outputs") return self._compute_loss(model, inputs) def _compute_loss(self, model, inputs): # Compute the per-token log probabilities for the model prompt_ids, prompt_mask = inputs["prompt_ids"], inputs["prompt_mask"] completion_ids, completion_mask = inputs["completion_ids"], inputs["completion_mask"] input_ids = torch.cat([prompt_ids, completion_ids], dim=1) attention_mask = torch.cat([prompt_mask, completion_mask], dim=1) logits_to_keep = completion_ids.size(1) # we only need to compute the logits for the completion tokens # Compute the per_token_logps and the entropy at each position in the completion per_token_logps, entropies = self._get_per_token_logps_and_entropies( model, input_ids, attention_mask, logits_to_keep, compute_entropy=True, ) logps = (per_token_logps * completion_mask).sum(1) # mask out padding and tokens after EOS old_logps = inputs["old_logps"] log_ratio = logps - old_logps # Compute the loss advantages = inputs["advantages"] coef_1 = torch.exp(log_ratio) coef_2 = torch.clamp(coef_1, 1 - self.epsilon_low, 1 + self.epsilon_high) per_sequence_loss1 = coef_1 * advantages per_sequence_loss2 = coef_2 * advantages per_sequence_loss = -torch.min(per_sequence_loss1, per_sequence_loss2) loss = per_sequence_loss.mean() # Log the metrics mode = "train" if self.model.training else "eval" # Entropy mean_entropy = (entropies * completion_mask).sum() / completion_mask.sum().clamp(min=1.0) self._metrics[mode]["entropy"].append(self.accelerator.gather(mean_entropy).nanmean().item()) # Compute the clipped probability ratios is_low_clipped = (coef_1 < 1 - self.epsilon_low) & (advantages < 0) is_high_clipped = (coef_1 > 1 + self.epsilon_high) & (advantages > 0) is_region_clipped = is_low_clipped | is_high_clipped gathered_low_clip = self.accelerator.gather(is_low_clipped.float().mean()) self._metrics[mode]["clip_ratio/low_mean"].append(gathered_low_clip.nanmean().item()) self._metrics[mode]["clip_ratio/low_min"].append(nanmin(gathered_low_clip).item()) gathered_high_clip = self.accelerator.gather(is_high_clipped.float().mean()) self._metrics[mode]["clip_ratio/high_mean"].append(gathered_high_clip.nanmean().item()) self._metrics[mode]["clip_ratio/high_max"].append(nanmax(gathered_high_clip).item()) gathered_clip_ratio = self.accelerator.gather(is_region_clipped.float().mean()) self._metrics[mode]["clip_ratio/region_mean"].append(gathered_clip_ratio.nanmean().item()) return loss def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys: Optional[list[str]] = None): inputs = self._prepare_inputs(inputs) with torch.no_grad(): with self.compute_loss_context_manager(): loss = self.compute_loss(model, inputs) loss = loss.mean().detach() return loss, None, None def log(self, logs: dict[str, float], start_time: Optional[float] = None) -> None: mode = "train" if self.model.training else "eval" metrics = {key: sum(val) / len(val) for key, val in self._metrics[mode].items()} # average the metrics # This method can be called both in training and evaluation. When called in evaluation, the keys in `logs` # start with "eval_". We need to add the prefix "eval_" to the keys in `metrics` to match the format. if mode == "eval": metrics = {f"eval_{key}": val for key, val in metrics.items()} logs = {**logs, **metrics} super().log(logs, start_time) self._metrics[mode].clear() if self.accelerator.is_main_process and self.log_completions: if is_rich_available(): print_prompt_completions_sample( self._logs["prompt"], self._logs["completion"], self._logs["rewards"], self._logs["advantages"], self.state.global_step, self.num_completions_to_print, ) if self.args.report_to and "wandb" in self.args.report_to and wandb.run is not None: import pandas as pd table = { "step": [str(self.state.global_step)] * len(self._logs["prompt"]), "prompt": self._logs["prompt"], "completion": self._logs["completion"], **self._logs["rewards"], "advantage": self._logs["advantages"], } df = pd.DataFrame(table) if self.wandb_log_unique_prompts: df = df.drop_duplicates(subset=["prompt"]) wandb.log({"completions": wandb.Table(dataframe=df)}) # Ensure the model card is saved along with the checkpoint def _save_checkpoint(self, model, trial): if self.args.hub_model_id is None: model_name = Path(self.args.output_dir).name else: model_name = self.args.hub_model_id.split("/")[-1] self.create_model_card(model_name=model_name) super()._save_checkpoint(model, trial) def create_model_card( self, model_name: Optional[str] = None, dataset_name: Optional[str] = None, tags: Union[str, list[str], None] = None, ): """ Creates a draft of a model card using the information available to the `Trainer`. Args: model_name (`str` or `None`, *optional*, defaults to `None`): Name of the model. dataset_name (`str` or `None`, *optional*, defaults to `None`): Name of the dataset used for training. tags (`str`, `list[str]` or `None`, *optional*, defaults to `None`): Tags to be associated with the model card. """ if not self.is_world_process_zero(): return if hasattr(self.model.config, "_name_or_path") and not os.path.isdir(self.model.config._name_or_path): base_model = self.model.config._name_or_path else: base_model = None # normalize `tags` to a mutable set if tags is None: tags = set() elif isinstance(tags, str): tags = {tags} else: tags = set(tags) if hasattr(self.model.config, "unsloth_version"): tags.add("unsloth") if "JOB_ID" in os.environ: tags.add("hf_jobs") tags.update(self._tag_names) # docstyle-ignore citation = textwrap.dedent( """\ @inproceedings{ahmadian2024back, title = {{Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs}}, author = {Arash Ahmadian and Chris Cremer and Matthias Gall{\'{e}} and Marzieh Fadaee and Julia Kreutzer and Olivier Pietquin and Ahmet {\"{U}}st{\"{u}}n and Sara Hooker}, year = 2024, booktitle = {Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), {ACL} 2024, Bangkok, Thailand, August 11-16, 2024}, pages = {12248--12267}, publisher = {Association for Computational Linguistics}, editor = {Lun{-}Wei Ku and Andre Martins and Vivek Srikumar}, } """ ) model_card = generate_model_card( base_model=base_model, model_name=model_name, hub_model_id=self.hub_model_id, dataset_name=dataset_name, tags=tags, wandb_url=wandb.run.url if is_wandb_available() and wandb.run is not None else None, comet_url=get_comet_experiment_url(), trainer_name="RLOO", trainer_citation=citation, paper_title="Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs", paper_id="2402.14740", ) model_card.save(os.path.join(self.args.output_dir, "README.md"))