main_spin.py

# Copyright 2024 Bytedance Ltd. and/or its affiliates
# Copyright 2023-2024 SGLang Team
#
# 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 os

import hydra
import ray

from recipe.spin.spin_trainer import RaySPINTrainer
from verl.trainer.ppo.reward import get_custom_reward_fn


@hydra.main(config_path="config", config_name="spin_trainer", version_base=None)
def main(config):
    run_ppo(config)


def run_ppo(config) -> None:
    # TODO(linjunrong.ocss884): this ENV is left for resolving SGLang conflict with ray devices
    # isolation, will solve in the future
    os.environ["ENSURE_CUDA_VISIBLE_DEVICES"] = os.environ.get("CUDA_VISIBLE_DEVICES", "")
    if not ray.is_initialized():
        # this is for local ray cluster
        ray.init(
            runtime_env={
                "env_vars": {"TOKENIZERS_PARALLELISM": "true", "NCCL_DEBUG": "WARN", "VLLM_LOGGING_LEVEL": "WARN"}
            }
        )

    runner = TaskRunner.remote()
    ray.get(runner.run.remote(config))


@ray.remote(num_cpus=1)  # please make sure main_task is not scheduled on head
class TaskRunner:
    def run(self, config):
        # print initial config
        from pprint import pprint

        from omegaconf import OmegaConf

        from verl.utils.fs import copy_to_local

        pprint(OmegaConf.to_container(config, resolve=True))  # resolve=True will eval symbol values
        OmegaConf.resolve(config)

        # download the checkpoint from hdfs
        local_path = copy_to_local(config.actor_rollout_ref.model.path)

        # instantiate tokenizer
        from verl.utils import hf_processor, hf_tokenizer

        trust_remote_code = config.data.get("trust_remote_code", False)
        tokenizer = hf_tokenizer(local_path, trust_remote_code=trust_remote_code)
        processor = hf_processor(local_path, use_fast=True)  # used for multimodal LLM, could be none

        # define worker classes
        if config.actor_rollout_ref.actor.strategy in {"fsdp", "fsdp2"}:
            assert config.critic.strategy in {"fsdp", "fsdp2"}
            # from recipe.spin.fsdp_workers import ActorRolloutRefWorker
            from recipe.spin.fsdp_workers import SPINRolloutRefWorker
            from verl.single_controller.ray import RayWorkerGroup

            ray_worker_group_cls = RayWorkerGroup

        elif config.actor_rollout_ref.actor.strategy == "megatron":
            assert config.actor_rollout_ref.actor.strategy == config.critic.strategy
            from verl.single_controller.ray.megatron import NVMegatronRayWorkerGroup

            ray_worker_group_cls = NVMegatronRayWorkerGroup

        else:
            raise NotImplementedError

        from recipe.spin.spin_trainer import ResourcePoolManager, Role

        role_worker_mapping = {
            # Role.ActorRollout: ray.remote(ActorRolloutRefWorker),
            Role.ActorRollout: ray.remote(SPINRolloutRefWorker),
            # Role.Critic: ray.remote(CriticWorker),
        }

        global_pool_id = "global_pool"
        resource_pool_spec = {
            global_pool_id: [config.trainer.n_gpus_per_node] * config.trainer.nnodes,
        }
        mapping = {
            Role.ActorRollout: global_pool_id,
            # Role.Critic: global_pool_id,
        }

        if config.reward_model.enable:
            if config.reward_model.strategy in {"fsdp", "fsdp2"}:
                from recipe.spin.fsdp_workers import RewardModelWorker
            elif config.reward_model.strategy == "megatron":
                from verl.workers.megatron_workers import RewardModelWorker
            else:
                raise NotImplementedError
            role_worker_mapping[Role.RewardModel] = ray.remote(RewardModelWorker)
            mapping[Role.RewardModel] = global_pool_id

        # use reference model
        # if config.algorithm.use_kl_in_reward or config.actor_rollout_ref.actor.use_kl_loss:
        # role_worker_mapping[Role.RefPolicy] = ray.remote(ActorRolloutRefWorker)
        role_worker_mapping[Role.RefPolicy] = ray.remote(SPINRolloutRefWorker)
        mapping[Role.RefPolicy] = global_pool_id

        from verl.workers.reward_manager import get_reward_manager_cls

        # Note(haibin.lin): please make sure custom reward managers are imported and
        # registered via `verl.workers.reward_manager.register`
        reward_manager_name = config.reward_model.get("reward_manager", "naive")
        reward_manager_cls = get_reward_manager_cls(reward_manager_name)

        compute_score = get_custom_reward_fn(config)
        reward_kwargs = dict(config.reward_model.get("reward_kwargs", {}))
        reward_fn = reward_manager_cls(
            tokenizer=tokenizer,
            num_examine=0,
            compute_score=compute_score,
            reward_fn_key=config.data.reward_fn_key,
            **reward_kwargs,
        )

        # Note that we always use function-based RM for validation
        val_reward_fn = reward_manager_cls(
            tokenizer=tokenizer, num_examine=1, compute_score=compute_score, reward_fn_key=config.data.reward_fn_key
        )
        resource_pool_manager = ResourcePoolManager(resource_pool_spec=resource_pool_spec, mapping=mapping)

        trainer = RaySPINTrainer(
            config=config,
            tokenizer=tokenizer,
            processor=processor,
            role_worker_mapping=role_worker_mapping,
            resource_pool_manager=resource_pool_manager,
            ray_worker_group_cls=ray_worker_group_cls,
            reward_fn=reward_fn,
            val_reward_fn=val_reward_fn,
        )
        trainer.init_workers()
        trainer.fit_dpo()


if __name__ == "__main__":
    main()