ray_trainer.py

# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# 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.
"""
FSDP PPO Trainer with Ray-based single controller.
This trainer supports model-agonistic model initialization with huggingface
"""

import os
import uuid
from contextlib import contextmanager
from copy import deepcopy
from dataclasses import dataclass, field
from enum import Enum
from pprint import pprint
from typing import Any, Dict, Optional, Type

import numpy as np
import torch
from codetiming import Timer
from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
from transformers import PreTrainedTokenizer, ProcessorMixin

from verl import DataProto
from verl.protocol import pad_dataproto_to_divisor, unpad_dataproto
from verl.single_controller.base import Worker
from verl.single_controller.ray import RayClassWithInitArgs, RayResourcePool, RayWorkerGroup
from verl.single_controller.ray.base import create_colocated_worker_cls
from verl.trainer import core_algos
from verl.trainer.config import PPOConfig
from verl.utils.rl_dataset import RLHFDataset, collate_fn
from verl.utils.torch_functional import masked_mean
from verl.utils.tracking import Tracking
from verl.workers.fsdp_workers import FSDPWorker


WorkerType = Type[Worker]


class Role(Enum):
    """
    To create more roles dynamically, you can subclass Role and add new members
    """

    Actor = 0
    Rollout = 1
    ActorRollout = 2
    Critic = 3
    RefPolicy = 4
    RewardModel = 5
    ActorRolloutRef = 6


@dataclass
class ResourcePoolManager:
    """
    Define a resource pool specification. Resource pool will be initialized first.
    Mapping
    """

    resource_pool_spec: dict[str, list[int]]
    mapping: dict[Role, str]
    resource_pool_dict: dict[str, RayResourcePool] = field(default_factory=dict)

    def create_resource_pool(self):
        for resource_pool_name, process_on_nodes in self.resource_pool_spec.items():
            # max_colocate_count means the number of WorkerGroups (i.e. processes) in each RayResourcePool
            # For FSDP backend, we recommend using max_colocate_count=1 that merge all WorkerGroups into one.
            # For Megatron backend, we recommend using max_colocate_count>1 that can utilize different WorkerGroup for differnt models
            resource_pool = RayResourcePool(
                process_on_nodes=process_on_nodes, use_gpu=True, max_colocate_count=1, name_prefix=resource_pool_name
            )
            self.resource_pool_dict[resource_pool_name] = resource_pool

    def get_resource_pool(self, role: Role) -> RayResourcePool:
        """Get the resource pool of the worker_cls"""
        return self.resource_pool_dict[self.mapping[role]]


def apply_kl_penalty(data: DataProto, kl_ctrl: core_algos.AdaptiveKLController, kl_penalty="kl"):
    responses = data.batch["responses"]
    response_length = responses.size(1)
    token_level_scores = data.batch["token_level_scores"]
    batch_size = data.batch.batch_size[0]
    attention_mask = data.batch["attention_mask"]
    response_mask = attention_mask[:, -response_length:]

    # compute kl between ref_policy and current policy
    if "ref_log_prob" in data.batch.keys():
        kld = core_algos.kl_penalty(
            data.batch["old_log_probs"], data.batch["ref_log_prob"], kl_penalty=kl_penalty
        )  # (batch_size, response_length)
        kld = kld * response_mask
        beta = kl_ctrl.value
    else:
        beta = 0
        kld = torch.zeros_like(response_mask, dtype=torch.float32)

    token_level_rewards = token_level_scores - beta * kld

    current_kl = masked_mean(kld, mask=response_mask, axis=-1)  # average over sequence
    current_kl = torch.mean(current_kl, dim=0).item()

    # according to https://github.com/huggingface/trl/blob/951ca1841f29114b969b57b26c7d3e80a39f75a0/trl/trainer/ppo_trainer.py#L837
    kl_ctrl.update(current_kl=current_kl, n_steps=batch_size)
    data.batch["token_level_rewards"] = token_level_rewards

    metrics = {"critic/kl": current_kl, "critic/kl_coeff": beta}

    return data, metrics


def compute_advantage(data: DataProto, adv_estimator, gamma=1.0, lam=1.0, num_repeat=1):
    # prepare response group
    # TODO: add other ways to estimate advantages
    if adv_estimator == "gae":
        values = data.batch["values"]
        responses = data.batch["responses"]
        response_length = responses.size(-1)
        attention_mask = data.batch["attention_mask"]
        response_mask = attention_mask[:, -response_length:]
        token_level_rewards = data.batch["token_level_rewards"]
        advantages, returns = core_algos.compute_gae_advantage_return(
            token_level_rewards=token_level_rewards, values=values, eos_mask=response_mask, gamma=gamma, lam=lam
        )
        data.batch["advantages"] = advantages
        data.batch["returns"] = returns
    elif adv_estimator == "grpo":
        token_level_rewards = data.batch["token_level_rewards"]
        index = data.non_tensor_batch["uid"]
        responses = data.batch["responses"]
        response_length = responses.size(-1)
        attention_mask = data.batch["attention_mask"]
        response_mask = attention_mask[:, -response_length:]
        advantages, returns = core_algos.compute_grpo_outcome_advantage(
            token_level_rewards=token_level_rewards, eos_mask=response_mask, index=index
        )
        data.batch["advantages"] = advantages
        data.batch["returns"] = returns
    elif adv_estimator == "reinforce_plus_plus":
        token_level_rewards = data.batch["token_level_rewards"]
        responses = data.batch["responses"]
        response_length = responses.size(-1)
        attention_mask = data.batch["attention_mask"]
        response_mask = attention_mask[:, -response_length:]
        advantages, returns = core_algos.compute_reinforce_plus_plus_outcome_advantage(
            token_level_rewards=token_level_rewards, eos_mask=response_mask, gamma=gamma
        )
        data.batch["advantages"] = advantages
        data.batch["returns"] = returns
    elif adv_estimator == "remax":
        token_level_rewards = data.batch["token_level_rewards"]
        index = data.non_tensor_batch["uid"]
        responses = data.batch["responses"]
        response_length = responses.size(-1)
        attention_mask = data.batch["attention_mask"]
        response_mask = attention_mask[:, -response_length:]

        reward_baselines = data.batch["reward_baselines"]

        advantages, returns = core_algos.compute_remax_outcome_advantage(
            token_level_rewards=token_level_rewards, reward_baselines=reward_baselines, eos_mask=response_mask
        )

        data.batch["advantages"] = advantages
        data.batch["returns"] = returns
    else:
        raise NotImplementedError
    return data


def reduce_metrics(metrics: Dict[str, Any]):
    for key, val in metrics.items():
        metrics[key] = np.mean(val)

    return metrics


def _compute_response_info(batch: DataProto):
    response_length = batch.batch["responses"].shape[-1]

    prompt_mask = batch.batch["attention_mask"][:, :-response_length]
    response_mask = batch.batch["attention_mask"][:, -response_length:]

    prompt_length = prompt_mask.sum(-1).float()
    response_length = response_mask.sum(-1).float()  # (batch_size,)

    return dict(
        response_mask=response_mask,
        prompt_length=prompt_length,
        response_length=response_length,
    )


def compute_data_metrics(batch: DataProto, use_critic: bool = True):
    # TODO: add response length
    sequence_score = batch.batch["token_level_scores"].sum(-1)
    sequence_reward = batch.batch["token_level_rewards"].sum(-1)

    advantages = batch.batch["advantages"]
    returns = batch.batch["returns"]

    max_response_length = batch.batch["responses"].shape[-1]

    prompt_mask = batch.batch["attention_mask"][:, :-max_response_length].bool()
    response_mask = batch.batch["attention_mask"][:, -max_response_length:].bool()

    max_prompt_length = prompt_mask.size(-1)

    response_info = _compute_response_info(batch)
    prompt_length = response_info["prompt_length"]
    response_length = response_info["response_length"]

    valid_adv = torch.masked_select(advantages, response_mask)
    valid_returns = torch.masked_select(returns, response_mask)

    if use_critic:
        values = batch.batch["values"]
        valid_values = torch.masked_select(values, response_mask)
        return_diff_var = torch.var(valid_returns - valid_values)
        return_var = torch.var(valid_returns)

    metrics = {
        # score
        "critic/score/mean": torch.mean(sequence_score).detach().item(),
        "critic/score/max": torch.max(sequence_score).detach().item(),
        "critic/score/min": torch.min(sequence_score).detach().item(),
        # reward
        "critic/rewards/mean": torch.mean(sequence_reward).detach().item(),
        "critic/rewards/max": torch.max(sequence_reward).detach().item(),
        "critic/rewards/min": torch.min(sequence_reward).detach().item(),
        # adv
        "critic/advantages/mean": torch.mean(valid_adv).detach().item(),
        "critic/advantages/max": torch.max(valid_adv).detach().item(),
        "critic/advantages/min": torch.min(valid_adv).detach().item(),
        # returns
        "critic/returns/mean": torch.mean(valid_returns).detach().item(),
        "critic/returns/max": torch.max(valid_returns).detach().item(),
        "critic/returns/min": torch.min(valid_returns).detach().item(),
        **(
            {
                # values
                "critic/values/mean": torch.mean(valid_values).detach().item(),
                "critic/values/max": torch.max(valid_values).detach().item(),
                "critic/values/min": torch.min(valid_values).detach().item(),
                # vf explained var
                "critic/vf_explained_var": (1.0 - return_diff_var / (return_var + 1e-5)).detach().item(),
            }
            if use_critic
            else {}
        ),
        # response length
        "response_length/mean": torch.mean(response_length).detach().item(),
        "response_length/max": torch.max(response_length).detach().item(),
        "response_length/min": torch.min(response_length).detach().item(),
        "response_length/clip_ratio": torch.mean(torch.eq(response_length, max_response_length).float())
        .detach()
        .item(),
        # prompt length
        "prompt_length/mean": torch.mean(prompt_length).detach().item(),
        "prompt_length/max": torch.max(prompt_length).detach().item(),
        "prompt_length/min": torch.min(prompt_length).detach().item(),
        "prompt_length/clip_ratio": torch.mean(torch.eq(prompt_length, max_prompt_length).float()).detach().item(),
    }
    return metrics


def compute_timing_metrics(batch, timing_raw):
    response_info = _compute_response_info(batch)
    num_prompt_tokens = torch.sum(response_info["prompt_length"]).item()
    num_response_tokens = torch.sum(response_info["response_length"]).item()
    num_overall_tokens = num_prompt_tokens + num_response_tokens

    num_tokens_of_section = {
        "gen": num_response_tokens,
        **{name: num_overall_tokens for name in ["ref", "values", "adv", "update_critic", "update_actor"]},
    }

    return {
        **{f"timing_s/{name}": value for name, value in timing_raw.items()},
        **{
            f"timing_per_token_ms/{name}": timing_raw[name] * 1000 / num_tokens_of_section[name]
            for name in set(num_tokens_of_section.keys()) & set(timing_raw.keys())
        },
    }


@contextmanager
def _timer(name: str, timing_raw: Dict[str, float]):
    with Timer(name=name, logger=None) as timer:
        yield

    timing_raw[name] = timer.last


class RayPPOTrainer:
    """
    Note that this trainer runs on the driver process on a single CPU/GPU node.
    """

    # TODO: support each role have individual ray_worker_group_cls,
    # i.e., support different backend of different role
    def __init__(
        self,
        config: PPOConfig,
        tokenizer: PreTrainedTokenizer,
        processor: Optional[ProcessorMixin],
        role_worker_mapping: dict[Role, WorkerType],
        resource_pool_manager: ResourcePoolManager,
        ray_worker_group_cls: RayWorkerGroup = RayWorkerGroup,
        reward_fn=None,
        val_reward_fn=None,
    ):
        self.tokenizer = tokenizer
        self.processor = processor
        self.config = config
        self.reward_fn = reward_fn
        self.val_reward_fn = val_reward_fn

        self.hybrid_engine = config.worker.hybrid_engine
        assert self.hybrid_engine, "Currently, only support hybrid engine"

        if self.hybrid_engine:
            assert Role.ActorRollout in role_worker_mapping, f"{role_worker_mapping.keys()}"

        self.role_worker_mapping = role_worker_mapping
        self.resource_pool_manager = resource_pool_manager
        self.use_reference_policy = Role.RefPolicy in role_worker_mapping
        self.use_reward_model = Role.RewardModel in role_worker_mapping
        self.ray_worker_group_cls = ray_worker_group_cls

        # define KL control
        if self.use_reference_policy:
            self.kl_ctrl = core_algos.get_kl_controller(config.algorithm)
        else:
            self.kl_ctrl = core_algos.FixedKLController(kl_coef=0.0)

        if self.config.algorithm.adv_estimator == "gae":
            self.use_critic = True
        elif self.config.algorithm.adv_estimator == "grpo":
            self.use_critic = False
        elif self.config.algorithm.adv_estimator == "reinforce_plus_plus":
            self.use_critic = False
        elif self.config.algorithm.adv_estimator == "remax":
            self.use_critic = False
        else:
            raise NotImplementedError

        self._create_dataloader()

    def _create_dataloader(self):
        self.train_dataset = RLHFDataset(
            data_path=self.config.data.train_files,
            tokenizer=self.tokenizer,
            processor=self.processor,
            prompt_key=self.config.data.prompt_key,
            max_prompt_length=self.config.data.max_prompt_length,
            truncation="right",
            system_prompt=self.config.data.system_prompt,
            min_pixels=self.config.data.min_pixels,
            max_pixels=self.config.data.max_pixels,
        )
        # use sampler for better ckpt resume
        if self.config.data.shuffle:
            train_dataloader_generator = torch.Generator()
            train_dataloader_generator.manual_seed(self.config.data.seed)
            sampler = RandomSampler(data_source=self.train_dataset, generator=train_dataloader_generator)
        else:
            sampler = SequentialSampler(data_source=self.train_dataset)

        self.train_dataloader = DataLoader(
            dataset=self.train_dataset,
            batch_size=self.config.data.rollout_batch_size,
            num_workers=8,
            drop_last=True,
            collate_fn=collate_fn,
            sampler=sampler,
        )

        self.val_dataset = RLHFDataset(
            data_path=self.config.data.val_files,
            tokenizer=self.tokenizer,
            processor=self.processor,
            prompt_key=self.config.data.prompt_key,
            max_prompt_length=self.config.data.max_prompt_length,
            truncation="right",
            system_prompt=self.config.data.system_prompt,
            min_pixels=self.config.data.min_pixels,
            max_pixels=self.config.data.max_pixels,
        )
        self.val_dataloader = DataLoader(
            dataset=self.val_dataset,
            batch_size=len(self.val_dataset),
            num_workers=8,
            shuffle=False,
            drop_last=False,
            collate_fn=collate_fn,
        )

        assert len(self.train_dataloader) >= 1
        assert len(self.val_dataloader) >= 1

        print(f"Size of train dataloader: {len(self.train_dataloader)}")
        print(f"Size of val dataloader: {len(self.val_dataloader)}")

        if self.config.trainer.max_steps is not None:
            training_steps = self.config.trainer.max_steps
        else:
            training_steps = len(self.train_dataloader) * self.config.trainer.total_episodes

        self.training_steps = training_steps
        self.config.worker.actor.optim.training_steps = training_steps
        self.config.worker.critic.optim.training_steps = training_steps
        print(f"Total training steps: {self.training_steps}")

    def _maybe_log_val_generations_to_wandb(self, inputs, outputs, scores):
        """Log a table of validation samples to wandb"""

        generations_to_log = self.config.trainer.val_generations_to_log_to_wandb

        if generations_to_log == 0:
            return

        if generations_to_log > 0 and "wandb" not in self.config.trainer.logger:
            print("WARNING: `val_generations_to_log_to_wandb` is set, but no wandb logger is found.")
            return

        import wandb

        # Create tuples of (input, output, score) and sort by input text
        samples = list(zip(inputs, outputs, scores))
        samples.sort(key=lambda x: x[0])  # Sort by input text

        # Use fixed random seed for deterministic shuffling
        rng = np.random.RandomState(42)
        rng.shuffle(samples)

        # Take first N samples after shuffling
        samples = samples[:generations_to_log]

        # Create column names for all samples
        columns = ["step"] + sum(
            [[f"input_{i + 1}", f"output_{i + 1}", f"score_{i + 1}"] for i in range(len(samples))], []
        )

        if not hasattr(self, "validation_table"):
            # Initialize the table on first call
            self.validation_table = wandb.Table(columns=columns)

        # Create a new table with same columns and existing data
        # Workaround for https://github.com/wandb/wandb/issues/2981#issuecomment-1997445737
        new_table = wandb.Table(columns=columns, data=self.validation_table.data)

        # Add new row with all data
        row_data = []
        row_data.append(self.global_steps)
        for sample in samples:
            row_data.extend(sample)

        new_table.add_data(*row_data)

        # Update reference and log
        wandb.log({"val/generations": new_table}, step=self.global_steps)
        self.validation_table = new_table

    def _validate(self):
        reward_tensor_lst = []
        data_source_lst = []

        # Lists to collect samples for the table
        sample_inputs = []
        sample_outputs = []
        sample_scores = []

        for test_data in self.val_dataloader:
            test_batch = DataProto.from_single_dict(test_data)
            # Store original inputs
            input_ids = test_batch.batch["input_ids"]
            input_texts = [self.tokenizer.decode(ids, skip_special_tokens=True) for ids in input_ids]
            sample_inputs.extend(input_texts)

            if "pixel_values" in test_batch.non_tensor_batch.keys():
                test_gen_batch = test_batch.pop(
                    batch_keys=["input_ids", "attention_mask", "position_ids"],
                    non_tensor_batch_keys=["pixel_values", "image_grid_thw", "raw_prompt_ids", "images"],
                )
            else:
                test_gen_batch = test_batch.pop(
                    batch_keys=["input_ids", "attention_mask", "position_ids"],
                    non_tensor_batch_keys=["raw_prompt_ids"],
                )

            test_gen_batch.meta_info = {"do_sample": False}

            # pad to be divisible by dp_size
            test_gen_batch_padded, pad_size = pad_dataproto_to_divisor(
                test_gen_batch, self.actor_rollout_wg.world_size
            )
            test_output_gen_batch_padded = self.actor_rollout_wg.generate_sequences(test_gen_batch_padded)
            # unpad
            test_output_gen_batch = unpad_dataproto(test_output_gen_batch_padded, pad_size=pad_size)
            print("validation generation end")

            # Store generated outputs
            output_ids = test_output_gen_batch.batch["responses"]
            output_texts = [self.tokenizer.decode(ids, skip_special_tokens=True) for ids in output_ids]
            sample_outputs.extend(output_texts)

            test_batch = test_batch.union(test_output_gen_batch)

            # evaluate using reward_function
            reward_tensor = self.val_reward_fn(test_batch)

            # Store scores
            scores = reward_tensor.sum(-1).cpu().tolist()
            sample_scores.extend(scores)

            reward_tensor_lst.append(reward_tensor)
            data_source_lst.append(
                test_batch.non_tensor_batch.get("data_source", ["unknown"] * reward_tensor.shape[0])
            )

        self._maybe_log_val_generations_to_wandb(inputs=sample_inputs, outputs=sample_outputs, scores=sample_scores)

        reward_tensor = torch.cat(reward_tensor_lst, dim=0).sum(-1).cpu()  # (batch_size,)
        data_sources = np.concatenate(data_source_lst, axis=0)

        # evaluate test_score based on data source
        data_source_reward = {}
        for i in range(reward_tensor.shape[0]):
            data_source = data_sources[i]
            if data_source not in data_source_reward:
                data_source_reward[data_source] = []
            data_source_reward[data_source].append(reward_tensor[i].item())

        metric_dict = {}
        for data_source, rewards in data_source_reward.items():
            metric_dict[f"val/test_score/{data_source}"] = np.mean(rewards)

        return metric_dict

    def init_workers(self):
        """Init resource pool and worker group"""
        self.resource_pool_manager.create_resource_pool()

        self.resource_pool_to_cls = {pool: {} for pool in self.resource_pool_manager.resource_pool_dict.values()}

        # create actor and rollout
        if self.hybrid_engine:
            resource_pool = self.resource_pool_manager.get_resource_pool(Role.ActorRollout)
            actor_rollout_cls = RayClassWithInitArgs(
                cls=self.role_worker_mapping[Role.ActorRollout], config=self.config.worker, role="actor_rollout"
            )
            self.resource_pool_to_cls[resource_pool]["actor_rollout"] = actor_rollout_cls
        else:
            raise NotImplementedError

        # create critic
        if self.use_critic:
            resource_pool = self.resource_pool_manager.get_resource_pool(Role.Critic)
            critic_cls = RayClassWithInitArgs(
                cls=self.role_worker_mapping[Role.Critic], config=self.config.worker, role="critic"
            )
            self.resource_pool_to_cls[resource_pool]["critic"] = critic_cls

        # create reference policy if needed
        if self.use_reference_policy:
            resource_pool = self.resource_pool_manager.get_resource_pool(Role.RefPolicy)
            ref_policy_cls = RayClassWithInitArgs(
                self.role_worker_mapping[Role.RefPolicy], config=self.config.worker, role="ref"
            )
            self.resource_pool_to_cls[resource_pool]["ref"] = ref_policy_cls

        # create a reward model if reward_fn is None
        if self.use_reward_model:
            # we create a RM here
            resource_pool = self.resource_pool_manager.get_resource_pool(Role.RewardModel)
            rm_cls = RayClassWithInitArgs(
                cls=self.role_worker_mapping[Role.RewardModel], config=self.config.worker, role="reward"
            )
            self.resource_pool_to_cls[resource_pool]["rm"] = rm_cls

        # initialize WorkerGroup
        # NOTE: if you want to use a different resource pool for each role, which can support different parallel size,
        # you should not use `create_colocated_worker_cls`. Instead, directly pass different resource pool to different worker groups.
        # See https://github.com/volcengine/verl/blob/master/examples/ray/tutorial.ipynb for more information.
        all_wg = {}
        self.wg_dicts = []
        for resource_pool, class_dict in self.resource_pool_to_cls.items():
            worker_dict_cls = create_colocated_worker_cls(class_dict=class_dict)
            wg_dict = self.ray_worker_group_cls(resource_pool=resource_pool, ray_cls_with_init=worker_dict_cls)
            spawn_wg = wg_dict.spawn(prefix_set=class_dict.keys())
            all_wg.update(spawn_wg)
            # keep the referece of WorkerDict to support ray >= 2.31. Ref: https://github.com/ray-project/ray/pull/45699
            self.wg_dicts.append(wg_dict)

        if self.use_critic:
            self.critic_wg: FSDPWorker = all_wg["critic"]
            self.critic_wg.init_model()

        if self.use_reference_policy:
            self.ref_policy_wg: FSDPWorker = all_wg["ref"]
            self.ref_policy_wg.init_model()

        if self.use_reward_model:
            self.rm_wg: FSDPWorker = all_wg["rm"]
            self.rm_wg.init_model()

        # we should create rollout at the end so that vllm can have a better estimation of kv cache memory
        self.actor_rollout_wg: FSDPWorker = all_wg["actor_rollout"]
        self.actor_rollout_wg.init_model()

    def _save_checkpoint(self):
        # path: {save_checkpoint_path}/global_step_{global_steps}/actor
        local_global_step_folder = os.path.join(
            self.config.trainer.save_checkpoint_path, f"global_step_{self.global_steps}"
        )
        actor_local_path = os.path.join(local_global_step_folder, "actor")

        self.actor_rollout_wg.save_checkpoint(
            actor_local_path,
            self.global_steps,
            remove_previous_ckpt=self.config.trainer.remove_previous_ckpt,
        )

        if self.use_critic:
            critic_local_path = os.path.join(local_global_step_folder, "critic")
            self.critic_wg.save_checkpoint(
                critic_local_path,
                self.global_steps,
                remove_previous_ckpt=self.config.trainer.remove_previous_ckpt,
            )

        local_latest_checkpointed_iteration = os.path.join(
            self.config.trainer.save_checkpoint_path, "latest_checkpointed_iteration.txt"
        )
        with open(local_latest_checkpointed_iteration, "w") as f:
            f.write(str(self.global_steps))

    def _load_checkpoint(self):
        if self.config.trainer.load_checkpoint_path is None:
            return

        print(f"Load from checkpoint: {self.config.trainer.load_checkpoint_path}")
        actor_path = os.path.join(self.config.trainer.load_checkpoint_path, "actor")
        critic_path = os.path.join(self.config.trainer.load_checkpoint_path, "critic")
        self.actor_rollout_wg.load_checkpoint(
            actor_path, del_local_after_load=self.config.trainer.del_local_ckpt_after_load
        )
        if self.use_critic:
            self.critic_wg.load_checkpoint(
                critic_path, del_local_after_load=self.config.trainer.del_local_ckpt_after_load
            )

    def fit(self):
        """
        The training loop of PPO.
        The driver process only need to call the compute functions of the worker group through RPC to construct the PPO dataflow.
        The light-weight advantage computation is done on the driver process.
        """
        logger = Tracking(
            project_name=self.config.trainer.project_name,
            experiment_name=self.config.trainer.experiment_name,
            default_backend=self.config.trainer.logger,
            config=self.config.to_dict(),
        )
        self.global_steps = 0

        # load checkpoint before doing anything
        self._load_checkpoint()

        # perform validation before training
        # currently, we only support validation using the reward_function.
        if self.val_reward_fn is not None and self.config.trainer.val_before_train:
            val_metrics = self._validate()
            pprint(f"Initial validation metrics: {val_metrics}")
            logger.log(data=val_metrics, step=self.global_steps)
            if self.config.trainer.val_only:
                return

        for _ in range(self.config.trainer.total_episodes):
            for batch_dict in self.train_dataloader:
                self.global_steps += 1
                if self.global_steps >= self.training_steps:
                    break

                metrics = {}
                timing_raw = {}

                batch: DataProto = DataProto.from_single_dict(batch_dict)

                # pop those keys for generation
                if "pixel_values" in batch.non_tensor_batch.keys():
                    gen_batch = batch.pop(
                        batch_keys=["input_ids", "attention_mask", "position_ids"],
                        non_tensor_batch_keys=["pixel_values", "image_grid_thw", "raw_prompt_ids", "images"],
                    )
                else:
                    gen_batch = batch.pop(
                        batch_keys=["input_ids", "attention_mask", "position_ids"],
                        non_tensor_batch_keys=["raw_prompt_ids"],
                    )

                with _timer("step", timing_raw):
                    # generate a batch
                    with _timer("gen", timing_raw):  # wg: worker group
                        gen_batch_output = self.actor_rollout_wg.generate_sequences(gen_batch)

                    if self.config.algorithm.adv_estimator == "remax":
                        with _timer("gen_max", timing_raw):
                            gen_baseline_batch = deepcopy(gen_batch)
                            gen_baseline_batch.meta_info["do_sample"] = False
                            gen_baseline_output = self.actor_rollout_wg.generate_sequences(gen_baseline_batch)

                            batch = batch.union(gen_baseline_output)
                            reward_baseline_tensor = self.reward_fn(batch)
                            reward_baseline_tensor = reward_baseline_tensor.sum(dim=-1)

                            batch.pop(batch_keys=list(gen_baseline_output.batch.keys()))

                            batch.batch["reward_baselines"] = reward_baseline_tensor

                            del gen_baseline_batch, gen_baseline_output

                    batch.non_tensor_batch["uid"] = np.array(
                        [str(uuid.uuid4()) for _ in range(len(batch.batch))], dtype=object
                    )
                    # repeat to align with repeated responses in rollout
                    batch = batch.repeat(repeat_times=self.config.worker.rollout.n, interleave=True)
                    batch = batch.union(gen_batch_output)

                    # balance the number of valid tokens on each dp rank.
                    # Note that this breaks the order of data inside the batch.
                    # Please take care when you implement group based adv computation such as GRPO and rloo
                    # self._balance_batch(batch, metrics=metrics) # TODO: re-enable balance batch

                    # compute global_valid tokens
                    batch.meta_info["global_token_num"] = torch.sum(batch.batch["attention_mask"], dim=-1).tolist()

                    # recompute old_log_probs
                    with _timer("old_log_prob", timing_raw):
                        old_log_prob = self.actor_rollout_wg.compute_log_prob(batch)
                        batch = batch.union(old_log_prob)

                    if self.use_reference_policy:
                        # compute reference log_prob
                        with _timer("ref", timing_raw):
                            ref_log_prob = self.ref_policy_wg.compute_ref_log_prob(batch)
                            batch = batch.union(ref_log_prob)

                    # compute values
                    if self.use_critic:
                        with _timer("values", timing_raw):
                            values = self.critic_wg.compute_values(batch)
                            batch = batch.union(values)

                    with _timer("adv", timing_raw):
                        # compute scores. Support both model and function-based.
                        # We first compute the scores using reward model. Then, we call reward_fn to combine
                        # the results from reward model and rule-based results.
                        if self.use_reward_model:
                            raise NotImplementedError

                        # we combine with rule-based rm
                        reward_tensor = self.reward_fn(batch)
                        batch.batch["token_level_scores"] = reward_tensor

                        # compute rewards. apply_kl_penalty if available
                        if not self.config.worker.actor.use_kl_loss:  # not grpo
                            batch, kl_metrics = apply_kl_penalty(
                                batch, kl_ctrl=self.kl_ctrl, kl_penalty=self.config.algorithm.kl_penalty
                            )
                            metrics.update(kl_metrics)
                        else:
                            batch.batch["token_level_rewards"] = batch.batch["token_level_scores"]

                        # compute advantages, executed on the driver process
                        batch = compute_advantage(
                            batch,
                            adv_estimator=self.config.algorithm.adv_estimator,
                            gamma=self.config.algorithm.gamma,
                            lam=self.config.algorithm.lam,
                            num_repeat=self.config.worker.rollout.n,
                        )

                    # update critic
                    if self.use_critic:
                        with _timer("update_critic", timing_raw):
                            critic_output = self.critic_wg.update_critic(batch)

                        critic_output_metrics = reduce_metrics(critic_output.meta_info["metrics"])
                        metrics.update(critic_output_metrics)

                    # implement critic warmup
                    if self.config.trainer.critic_warmup <= self.global_steps:
                        # update actor
                        with _timer("update_actor", timing_raw):
                            actor_output = self.actor_rollout_wg.update_actor(batch)

                        actor_output_metrics = reduce_metrics(actor_output.meta_info["metrics"])
                        metrics.update(actor_output_metrics)

                    # validate
                    if (
                        self.val_reward_fn is not None
                        and self.config.trainer.test_freq > 0
                        and self.global_steps % self.config.trainer.test_freq == 0
                    ):
                        with _timer("testing", timing_raw):
                            val_metrics: dict = self._validate()
                        metrics.update(val_metrics)

                    if self.config.trainer.save_freq > 0 and self.global_steps % self.config.trainer.save_freq == 0:
                        with _timer("save_checkpoint", timing_raw):
                            self._save_checkpoint()

                # collect metrics
                metrics.update(compute_data_metrics(batch=batch, use_critic=self.use_critic))
                metrics.update(compute_timing_metrics(batch=batch, timing_raw=timing_raw))

                # TODO: make a canonical logger that supports various backend
                logger.log(data=metrics, step=self.global_steps)

        # perform validation after training
        if self.val_reward_fn is not None:
            val_metrics = self._validate()
            pprint(f"Final validation metrics: {val_metrics}")
            logger.log(data=val_metrics, step=self.global_steps)

        self._save_checkpoint()