trainers.py

import torch
torch.backends.cuda.matmul.allow_tf32 = True
import torch.nn.functional as F
import torch.nn as nn
import transformers
from omegaconf import DictConfig

import torch.distributed as dist
from torch.distributed.fsdp import (
    FullyShardedDataParallel as FSDP,
    MixedPrecision,
    StateDictType,
    BackwardPrefetch,
    ShardingStrategy,
    CPUOffload,
)
from torch.distributed.fsdp.api import FullStateDictConfig, FullOptimStateDictConfig
from torch.distributed.fsdp.wrap import transformer_auto_wrap_policy
import tensor_parallel as tp
import contextlib

from preference_datasets import get_batch_iterator
from utils import (
    slice_and_move_batch_for_device,
    formatted_dict,
    all_gather_if_needed,
    pad_to_length,
    get_block_class_from_model,
    rank0_print,
    get_local_dir,
)
import numpy as np
import wandb
import tqdm

import random
import os
from collections import defaultdict
import time
import json
import functools
from typing import Optional, Dict, List, Union, Tuple


def dpo_loss(policy_chosen_logps: torch.FloatTensor,
             policy_rejected_logps: torch.FloatTensor,
             reference_chosen_logps: torch.FloatTensor,
             reference_rejected_logps: torch.FloatTensor,
             beta: float,
             reference_free: bool = False) -> Tuple[torch.FloatTensor, torch.FloatTensor, torch.FloatTensor]:
    """Compute the DPO loss for a batch of policy and reference model log probabilities.
    
    Args:
        policy_chosen_logps: Log probabilities of the policy model for the chosen responses. Shape: (batch_size,)
        policy_rejected_logps: Log probabilities of the policy model for the rejected responses. Shape: (batch_size,)
        reference_chosen_logps: Log probabilities of the reference model for the chosen responses. Shape: (batch_size,)
        reference_rejected_logps: Log probabilities of the reference model for the rejected responses. Shape: (batch_size,)
        beta: Temperature parameter for the DPO loss, typically something in the range of 0.1 to 0.5. We ignore the reference model as beta -> 0.
        reference_free: If True, we ignore the _provided_ reference model and implicitly use a reference model that assigns equal probability to all responses.

    Returns:
        A tuple of three tensors: (losses, chosen_rewards, rejected_rewards).
        The losses tensor contains the DPO loss for each example in the batch.
        The chosen_rewards and rejected_rewards tensors contain the rewards for the chosen and rejected responses, respectively.
    """
    pi_logratios = policy_chosen_logps - policy_rejected_logps
    ref_logratios = reference_chosen_logps - reference_rejected_logps

    if reference_free:
        ref_logratios = 0

    logits = pi_logratios - ref_logratios

    losses = -F.logsigmoid(beta * logits)
    chosen_rewards = beta * (policy_chosen_logps - reference_chosen_logps).detach()
    rejected_rewards = beta * (policy_rejected_logps - reference_rejected_logps).detach()

    return losses, chosen_rewards, rejected_rewards


def _get_batch_logps(logits: torch.FloatTensor, labels: torch.LongTensor, average_log_prob: bool = False) -> torch.FloatTensor:
    """Compute the log probabilities of the given labels under the given logits.

    Args:
        logits: Logits of the model (unnormalized). Shape: (batch_size, sequence_length, vocab_size)
        labels: Labels for which to compute the log probabilities. Label tokens with a value of -100 are ignored. Shape: (batch_size, sequence_length)
        average_log_prob: If True, return the average log probability per (non-masked) token. Otherwise, return the sum of the log probabilities of the (non-masked) tokens.

    Returns:
        A tensor of shape (batch_size,) containing the average/sum log probabilities of the given labels under the given logits.
    """
    assert logits.shape[:-1] == labels.shape

    labels = labels[:, 1:].clone()
    logits = logits[:, :-1, :]
    loss_mask = (labels != -100)

    # dummy token; we'll ignore the losses on these tokens later
    labels[labels == -100] = 0

    per_token_logps = torch.gather(logits.log_softmax(-1), dim=2, index=labels.unsqueeze(2)).squeeze(2)

    if average_log_prob:
        return (per_token_logps * loss_mask).sum(-1) / loss_mask.sum(-1)
    else:
        return (per_token_logps * loss_mask).sum(-1)


def concatenated_inputs(batch: Dict[str, Union[List, torch.LongTensor]]) -> Dict[str, torch.LongTensor]:
    """Concatenate the chosen and rejected inputs into a single tensor.
    
    Args:
        batch: A batch of data. Must contain the keys 'chosen_input_ids' and 'rejected_input_ids', which are tensors of shape (batch_size, sequence_length).
        
    Returns:
        A dictionary containing the concatenated inputs under the key 'concatenated_input_ids'.
    """
    max_length = max(batch['chosen_input_ids'].shape[1], batch['rejected_input_ids'].shape[1])
    concatenated_batch = {}
    for k in batch:
        if k.startswith('chosen') and isinstance(batch[k], torch.Tensor):
            pad_value = -100 if 'labels' in k else 0
            concatenated_key = k.replace('chosen', 'concatenated')
            concatenated_batch[concatenated_key] = pad_to_length(batch[k], max_length, pad_value=pad_value)
    for k in batch:
        if k.startswith('rejected') and isinstance(batch[k], torch.Tensor):
            pad_value = -100 if 'labels' in k else 0
            concatenated_key = k.replace('rejected', 'concatenated')
            concatenated_batch[concatenated_key] = torch.cat((
                concatenated_batch[concatenated_key],
                pad_to_length(batch[k], max_length, pad_value=pad_value),
            ), dim=0)
    return concatenated_batch


class BasicTrainer(object):
    def __init__(self, policy: nn.Module, config: DictConfig, seed: int, run_dir: str, reference_model: Optional[nn.Module] = None, rank: int = 0, world_size: int = 1):
        """A trainer for a language model, supporting either SFT or DPO training.
           
           If multiple GPUs are present, naively splits the model across them, effectively
           offering N times available memory, but without any parallel computation.
        """
        self.seed = seed
        self.rank = rank
        self.world_size = world_size
        self.config = config
        self.run_dir = run_dir

        tokenizer_name_or_path = config.model.tokenizer_name_or_path or config.model.name_or_path
        rank0_print(f'Loading tokenizer {tokenizer_name_or_path}')
        self.tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer_name_or_path, cache_dir=get_local_dir(config.local_dirs))
        if self.tokenizer.pad_token_id is None:
            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id

        data_iterator_kwargs = dict(
            names=config.datasets,
            tokenizer=self.tokenizer,
            shuffle=True,
            max_length=config.max_length,
            max_prompt_length=config.max_prompt_length,
            sft_mode=config.loss.name == 'sft',
            prompt_before=config.prompt_before.replace('\\n', '\n'), # NOTE: prevent escaping \n (but currently not other special characters)
            prompt_after=config.prompt_after.replace('\\n', '\n'),
        )

        self.policy = policy
        self.reference_model = reference_model

        self.train_iterator = get_batch_iterator(**data_iterator_kwargs, split='train', n_epochs=config.n_epochs, n_examples=config.n_examples, batch_size=config.batch_size, silent=rank != 0, cache_dir=get_local_dir(config.local_dirs))
        rank0_print(f'Loaded train data iterator')
        self.eval_iterator = get_batch_iterator(**data_iterator_kwargs, split='test', n_examples=config.n_eval_examples, batch_size=config.eval_batch_size, silent=rank != 0, cache_dir=get_local_dir(config.local_dirs))
        self.eval_batches = list(self.eval_iterator)
        rank0_print(f'Loaded {len(self.eval_batches)} eval batches of size {config.eval_batch_size}')

    def get_batch_samples(self, batch: Dict[str, torch.LongTensor]) -> Tuple[str, str]:
        """Generate samples from the policy (and reference model, if doing DPO training) for the given batch of inputs."""

        # FSDP generation according to https://github.com/pytorch/pytorch/issues/100069
        ctx = lambda: (FSDP.summon_full_params(self.policy, writeback=False, recurse=False) if 'FSDP' in self.config.trainer else contextlib.nullcontext())
        with ctx():
            policy_output = self.policy.generate(
                batch['prompt_input_ids'], attention_mask=batch['prompt_attention_mask'], max_length=self.config.max_length, do_sample=True, pad_token_id=self.tokenizer.pad_token_id)

        if self.config.loss.name == 'dpo':
            ctx = lambda: (FSDP.summon_full_params(self.reference_model, writeback=False, recurse=False) if 'FSDP' in self.config.trainer else contextlib.nullcontext())
            with ctx():
                reference_output = self.reference_model.generate(
                    batch['prompt_input_ids'], attention_mask=batch['prompt_attention_mask'], max_length=self.config.max_length, do_sample=True, pad_token_id=self.tokenizer.pad_token_id)

        policy_output = pad_to_length(policy_output, self.config.max_length, self.tokenizer.pad_token_id)
        policy_output = all_gather_if_needed(policy_output, self.rank, self.world_size)
        policy_output_decoded = self.tokenizer.batch_decode(policy_output, skip_special_tokens=True)

        if self.config.loss.name == 'dpo':
            reference_output = pad_to_length(reference_output, self.config.max_length, self.tokenizer.pad_token_id)
            reference_output = all_gather_if_needed(reference_output, self.rank, self.world_size)
            reference_output_decoded = self.tokenizer.batch_decode(reference_output, skip_special_tokens=True)
        else:
            reference_output_decoded = []

        return policy_output_decoded, reference_output_decoded
    
    def concatenated_forward(self, model: nn.Module, batch: Dict[str, Union[List, torch.LongTensor]]) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
        """Run the given model on the given batch of inputs, concatenating the chosen and rejected inputs together.
        
           We do this to avoid doing two forward passes, because it's faster for FSDP.
        """
        concatenated_batch = concatenated_inputs(batch)
        all_logits = model(concatenated_batch['concatenated_input_ids'], attention_mask=concatenated_batch['concatenated_attention_mask']).logits.to(torch.float32)
        all_logps = _get_batch_logps(all_logits, concatenated_batch['concatenated_labels'], average_log_prob=False)
        chosen_logps = all_logps[:batch['chosen_input_ids'].shape[0]]
        rejected_logps = all_logps[batch['chosen_input_ids'].shape[0]:]
        return chosen_logps, rejected_logps


    def get_batch_metrics(self, batch: Dict[str, Union[List, torch.LongTensor]], loss_config: DictConfig, train=True):
        """Compute the SFT or DPO loss and other metrics for the given batch of inputs."""

        metrics = {}
        train_test = 'train' if train else 'eval'

        if loss_config.name == 'dpo':
            policy_chosen_logps, policy_rejected_logps = self.concatenated_forward(self.policy, batch)
            with torch.no_grad():
                reference_chosen_logps, reference_rejected_logps = self.concatenated_forward(self.reference_model, batch)

            losses, chosen_rewards, rejected_rewards = dpo_loss(
                policy_chosen_logps, policy_rejected_logps, reference_chosen_logps, reference_rejected_logps, beta=loss_config.beta, reference_free=loss_config.reference_free)
            reward_accuracies = (chosen_rewards > rejected_rewards).float()

            chosen_rewards = all_gather_if_needed(chosen_rewards, self.rank, self.world_size)
            rejected_rewards = all_gather_if_needed(rejected_rewards, self.rank, self.world_size)
            reward_accuracies = all_gather_if_needed(reward_accuracies, self.rank, self.world_size)

            metrics[f'rewards_{train_test}/chosen'] = chosen_rewards.cpu().numpy().tolist()
            metrics[f'rewards_{train_test}/rejected'] = rejected_rewards.cpu().numpy().tolist()
            metrics[f'rewards_{train_test}/accuracies'] = reward_accuracies.cpu().numpy().tolist()
            metrics[f'rewards_{train_test}/margins'] = (chosen_rewards - rejected_rewards).cpu().numpy().tolist()
            metrics[f'logps_{train_test}/accuracies'] = (policy_chosen_logps > policy_rejected_logps).cpu().numpy().tolist()

            policy_rejected_logps = all_gather_if_needed(policy_rejected_logps.detach(), self.rank, self.world_size)
            metrics[f'logps_{train_test}/rejected'] = policy_rejected_logps.cpu().numpy().tolist()

        elif loss_config.name == 'sft':
            policy_chosen_logits = self.policy(batch['chosen_input_ids'], attention_mask=batch['chosen_attention_mask']).logits.to(torch.float32)
            policy_chosen_logps = _get_batch_logps(policy_chosen_logits, batch['chosen_labels'], average_log_prob=False)

            losses = -policy_chosen_logps

        policy_chosen_logps = all_gather_if_needed(policy_chosen_logps.detach(), self.rank, self.world_size)
        metrics[f'logps_{train_test}/chosen'] = policy_chosen_logps.cpu().numpy().tolist()

        all_devices_losses = all_gather_if_needed(losses.detach(), self.rank, self.world_size)
        metrics[f'loss/{train_test}'] = all_devices_losses.cpu().numpy().tolist()

        return losses.mean(), metrics

    def train(self):
        """Begin either SFT or DPO training, with periodic evaluation."""

        rank0_print(f'Using {self.config.optimizer} optimizer')
        self.optimizer = getattr(torch.optim, self.config.optimizer)(self.policy.parameters(), lr=self.config.lr)
        self.scheduler = torch.optim.lr_scheduler.LambdaLR(self.optimizer, lr_lambda=lambda step: min(1.0, (step + 1) / (self.config.warmup_steps + 1)))
    
        torch.manual_seed(self.seed)
        np.random.seed(self.seed)
        random.seed(self.seed)

        if self.config.loss.name == 'dpo':
            self.reference_model.eval()

        self.example_counter = 0
        self.batch_counter = 0
        last_log = None

        for batch in self.train_iterator:
            #### BEGIN EVALUATION ####
            if self.example_counter % self.config.eval_every == 0 and (self.example_counter > 0 or self.config.do_first_eval):
                rank0_print(f'Running evaluation after {self.example_counter} train examples')
                self.policy.eval()

                all_eval_metrics = defaultdict(list)
                if self.config.sample_during_eval:
                    all_policy_samples, all_reference_samples = [], []
                    policy_text_table = wandb.Table(columns=["step", "prompt", "sample"])
                    if self.config.loss.name == 'dpo':
                        reference_text_table = wandb.Table(columns=["step", "prompt", "sample"])

                for eval_batch in (tqdm.tqdm(self.eval_batches, desc='Computing eval metrics') if self.rank == 0 else self.eval_batches):
                    local_eval_batch = slice_and_move_batch_for_device(eval_batch, self.rank, self.world_size, self.rank)
                    with torch.no_grad():
                        _, eval_metrics = self.get_batch_metrics(local_eval_batch, self.config.loss, train=False)

                    for k, v in eval_metrics.items():
                        all_eval_metrics[k].extend(v)

                if self.config.sample_during_eval:
                    if self.config.n_eval_model_samples < self.config.eval_batch_size:
                        rank0_print(f'Warning: n_eval_model_samples ({self.config.n_eval_model_samples}) < eval_batch_size ({self.config.eval_batch_size}). Sampling from the first complete eval batch of prompts.')
                        sample_batches = self.eval_batches[:1]
                    else:
                        n_sample_batches = self.config.n_eval_model_samples // self.config.eval_batch_size
                        sample_batches = self.eval_batches[:n_sample_batches]
                    for eval_batch in (tqdm.tqdm(sample_batches, desc='Generating samples...') if self.rank == 0 else sample_batches):
                        local_eval_batch = slice_and_move_batch_for_device(eval_batch, self.rank, self.world_size, self.rank)
                        policy_samples, reference_samples = self.get_batch_samples(local_eval_batch)

                        all_policy_samples.extend(policy_samples)
                        all_reference_samples.extend(reference_samples)

                        for prompt, sample in zip(eval_batch['prompt'], policy_samples):
                            policy_text_table.add_data(self.example_counter, prompt, sample)
                        if self.config.loss.name == 'dpo':
                            for prompt, sample in zip(eval_batch['prompt'], reference_samples):
                                reference_text_table.add_data(self.example_counter, prompt, sample)

                mean_eval_metrics = {k: sum(v) / len(v) for k, v in all_eval_metrics.items()}
                rank0_print(f'eval after {self.example_counter}: {formatted_dict(mean_eval_metrics)}')
                if self.config.sample_during_eval:                    
                    rank0_print(json.dumps(all_policy_samples[:10], indent=2))
                    if self.config.loss.name == 'dpo':
                        rank0_print(json.dumps(all_reference_samples[:10], indent=2))

                if self.config.wandb.enabled and self.rank == 0:
                    wandb.log(mean_eval_metrics, step=self.example_counter)

                    if self.config.sample_during_eval:
                        wandb.log({"policy_samples": policy_text_table}, step=self.example_counter)
                        if self.config.loss.name == 'dpo':
                            wandb.log({"reference_samples": reference_text_table}, step=self.example_counter)

                if self.example_counter > 0:
                    if self.config.debug:
                        rank0_print('skipping save in debug mode')
                    else:
                        output_dir = os.path.join(self.run_dir, f'step-{self.example_counter}')
                        rank0_print(f'creating checkpoint to write to {output_dir}...')
                        self.save(output_dir, mean_eval_metrics)
            #### END EVALUATION ####

            #### BEGIN TRAINING ####
            self.policy.train()

            start_time = time.time()
            batch_metrics = defaultdict(list)
            for microbatch_idx in range(self.config.gradient_accumulation_steps):
                global_microbatch = slice_and_move_batch_for_device(batch, microbatch_idx, self.config.gradient_accumulation_steps, self.rank)
                local_microbatch = slice_and_move_batch_for_device(global_microbatch, self.rank, self.world_size, self.rank)
                loss, metrics = self.get_batch_metrics(local_microbatch, self.config.loss, train=True)
                (loss / self.config.gradient_accumulation_steps).backward()

                for k, v in metrics.items():
                    batch_metrics[k].extend(v)

            grad_norm = self.clip_gradient()
            self.optimizer.step()
            self.scheduler.step()
            self.optimizer.zero_grad()

            step_time = time.time() - start_time
            examples_per_second = self.config.batch_size / step_time
            batch_metrics['examples_per_second'].append(examples_per_second)
            batch_metrics['grad_norm'].append(grad_norm)

            self.batch_counter += 1
            self.example_counter += self.config.batch_size

            if last_log is None or time.time() - last_log > self.config.minimum_log_interval_secs:
                mean_train_metrics = {k: sum(v) / len(v) for k, v in batch_metrics.items()}
                mean_train_metrics['counters/examples'] = self.example_counter
                mean_train_metrics['counters/updates'] = self.batch_counter
                rank0_print(f'train stats after {self.example_counter} examples: {formatted_dict(mean_train_metrics)}')

                if self.config.wandb.enabled and self.rank == 0:
                    wandb.log(mean_train_metrics, step=self.example_counter)

                last_log = time.time()
            else:
                rank0_print(f'skipping logging after {self.example_counter} examples to avoid logging too frequently')
            #### END TRAINING ####


    def clip_gradient(self):
        """Clip the gradient norm of the parameters of a non-FSDP policy."""
        return torch.nn.utils.clip_grad_norm_(self.policy.parameters(), self.config.max_grad_norm).item()

    def write_state_dict(self, step: int, state: Dict[str, torch.Tensor], metrics: Dict, filename: str, dir_name: Optional[str] = None):
        """Write a checkpoint to disk."""
        if dir_name is None:
            dir_name = os.path.join(self.run_dir, f'LATEST')

        os.makedirs(dir_name, exist_ok=True)
        output_path = os.path.join(dir_name, filename)
        rank0_print(f'writing checkpoint to {output_path}...')
        torch.save({
            'step_idx': step,
            'state': state,
            'metrics': metrics if metrics is not None else {},
        }, output_path)
    
    def save(self, output_dir: Optional[str] = None, metrics: Optional[Dict] = None):
        """Save policy, optimizer, and scheduler state to disk."""

        policy_state_dict = self.policy.state_dict()
        self.write_state_dict(self.example_counter, policy_state_dict, metrics, 'policy.pt', output_dir)
        del policy_state_dict

        optimizer_state_dict = self.optimizer.state_dict()
        self.write_state_dict(self.example_counter, optimizer_state_dict, metrics, 'optimizer.pt', output_dir)
        del optimizer_state_dict

        scheduler_state_dict = self.scheduler.state_dict()
        self.write_state_dict(self.example_counter, scheduler_state_dict, metrics, 'scheduler.pt', output_dir)


class FSDPTrainer(BasicTrainer):
    def __init__(self, policy: nn.Module, config: DictConfig, seed: int, run_dir: str, reference_model: Optional[nn.Module] = None, rank: int = 0, world_size: int = 1):
        """A trainer subclass that uses PyTorch FSDP to shard the model across multiple GPUs.
        
           This trainer will shard both the policy and reference model across all available GPUs.
           Models are sharded at the block level, where the block class name is provided in the config.
        """

        super().__init__(policy, config, seed, run_dir, reference_model, rank, world_size)
        assert config.model.block_name is not None, 'must specify model.block_name (e.g., GPT2Block or GPTNeoXLayer) for FSDP'

        wrap_class = get_block_class_from_model(policy, config.model.block_name)
        model_auto_wrap_policy = functools.partial(transformer_auto_wrap_policy, transformer_layer_cls={wrap_class},)

        shared_fsdp_kwargs = dict(
            auto_wrap_policy=model_auto_wrap_policy,
            sharding_strategy=ShardingStrategy.FULL_SHARD,
            cpu_offload=CPUOffload(offload_params=False),
            backward_prefetch=BackwardPrefetch.BACKWARD_PRE,
            device_id=rank,
            ignored_modules=None,
            limit_all_gathers=False,
            use_orig_params=False,
            sync_module_states=False
        )

        rank0_print('Sharding policy...')
        mp_dtype = getattr(torch, config.model.fsdp_policy_mp) if config.model.fsdp_policy_mp is not None else None
        policy_mp_policy = MixedPrecision(param_dtype=mp_dtype, reduce_dtype=mp_dtype, buffer_dtype=mp_dtype)
        self.policy = FSDP(policy, **shared_fsdp_kwargs, mixed_precision=policy_mp_policy)

        if config.activation_checkpointing:
            rank0_print('Attempting to enable activation checkpointing...')
            try:
                # use activation checkpointing, according to:
                # https://pytorch.org/blog/scaling-multimodal-foundation-models-in-torchmultimodal-with-pytorch-distributed/
                #
                # first, verify we have FSDP activation support ready by importing:
                from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
                    checkpoint_wrapper,
                    apply_activation_checkpointing,
                    CheckpointImpl,
                )
                non_reentrant_wrapper = functools.partial(
                    checkpoint_wrapper,
                    offload_to_cpu=False,
                    checkpoint_impl=CheckpointImpl.NO_REENTRANT,
                )
            except Exception as e:
                rank0_print('FSDP activation checkpointing not available:', e)
            else:
                check_fn = lambda submodule: isinstance(submodule, wrap_class)
                rank0_print('Applying activation checkpointing wrapper to policy...')
                apply_activation_checkpointing(self.policy, checkpoint_wrapper_fn=non_reentrant_wrapper, check_fn=check_fn)
                rank0_print('FSDP activation checkpointing enabled!')

        if config.loss.name == 'dpo':
            rank0_print('Sharding reference model...')
            self.reference_model = FSDP(reference_model, **shared_fsdp_kwargs)
        
        print('Loaded model on rank', rank)
        dist.barrier()

    def clip_gradient(self):
        """Clip the gradient norm of the parameters of an FSDP policy, gathering the gradients across all GPUs."""
        return self.policy.clip_grad_norm_(self.config.max_grad_norm).item()
    
    def save(self, output_dir=None, metrics=None):
        """Save policy, optimizer, and scheduler state to disk, gathering from all processes and saving only on the rank 0 process."""
        save_policy = FullStateDictConfig(offload_to_cpu=True, rank0_only=True)
        with FSDP.state_dict_type(self.policy, StateDictType.FULL_STATE_DICT, state_dict_config=save_policy):
            policy_state_dict = self.policy.state_dict()

        if self.rank == 0:
            self.write_state_dict(self.example_counter, policy_state_dict, metrics, 'policy.pt', output_dir)
        del policy_state_dict
        dist.barrier()

        save_policy = FullOptimStateDictConfig(offload_to_cpu=True, rank0_only=True)
        with FSDP.state_dict_type(self.policy, StateDictType.FULL_STATE_DICT, optim_state_dict_config=save_policy):
            optimizer_state_dict = FSDP.optim_state_dict(self.policy, self.optimizer)

        if self.rank == 0:
            self.write_state_dict(self.example_counter, optimizer_state_dict, metrics, 'optimizer.pt', output_dir)
        del optimizer_state_dict
        dist.barrier()

        if self.rank == 0:
            scheduler_state_dict = self.scheduler.state_dict()
            self.write_state_dict(self.example_counter, scheduler_state_dict, metrics, 'scheduler.pt', output_dir)
        dist.barrier()
        

class TensorParallelTrainer(BasicTrainer):
    def __init__(self, policy, config, seed, run_dir, reference_model=None, rank=0, world_size=1):
        """A trainer subclass that uses TensorParallel to shard the model across multiple GPUs.

           Based on https://github.com/BlackSamorez/tensor_parallel. Note sampling is extremely slow,
              see https://github.com/BlackSamorez/tensor_parallel/issues/66.
        """
        super().__init__(policy, config, seed, run_dir, reference_model, rank, world_size)
        
        rank0_print('Sharding policy...')
        self.policy = tp.tensor_parallel(policy, sharded=True)
        if config.loss.name == 'dpo':
            rank0_print('Sharding reference model...')
            self.reference_model = tp.tensor_parallel(reference_model, sharded=False)

    def save(self, output_dir=None, metrics=None):
        """Save (unsharded) policy state to disk."""
        with tp.save_tensor_parallel(self.policy):
            policy_state_dict = self.policy.state_dict()
    
        self.write_state_dict(self.example_counter, policy_state_dict, metrics, 'policy.pt', output_dir)
        del policy_state_dict