Function main_task

verl/examples/split_placement/main_ppo_split.py:111–213

(config)

@ray.remote
def main_task(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_tokenizer

    tokenizer = hf_tokenizer(local_path)

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

        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 import RayWorkerGroup
        from verl.workers.megatron_workers import ActorRolloutRefWorker, CriticWorker

        ray_worker_group_cls = RayWorkerGroup

    else:
        raise NotImplementedError

    from verl.trainer.ppo.ray_trainer import ResourcePoolManager, Role

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

    # NOTE: initialze two resource pool
    actor_rollout_ref_pool_id = "actor_rollout_ref_pool"
    critic_pool_id = "critic_pool"
    if config.trainer.nnodes // 2 == 0 and config.trainer.n_gpus_per_node // 2 > 0:
        resource_pool_spec = {
            actor_rollout_ref_pool_id: [config.trainer.n_gpus_per_node // 2] * config.trainer.nnodes,
            critic_pool_id: [config.trainer.n_gpus_per_node // 2] * config.trainer.nnodes,
        }
    else:
        resource_pool_spec = {
            actor_rollout_ref_pool_id: [config.trainer.n_gpus_per_node] * (config.trainer.nnodes // 2),
            critic_pool_id: [config.trainer.n_gpus_per_node] * (config.trainer.nnodes // 2),
        }
    print(f"resource_pool_spec: {resource_pool_spec}")