trinity.trainer.trainer module

trinity.trainer.trainer module#

Trainer Class

class trinity.trainer.trainer.Trainer(config: Config)[source]#

Bases: object

Consume the experience and train the model.

__init__(config: Config) → None[source]#

async prepare() → None[source]#: Prepare the trainer.

async get_weight_sync_info() → Tuple[str, int, List] | None[source]#

Get rendezvous info for NCCL weight sync group setup.

Returns (master_address, master_port, state_dict_meta) from the trainer’s GPU worker rank 0. Called by Synchronizer before coordinating NCCL group creation.

async setup_weight_sync_group(master_address: str, master_port: int, world_size: int, group_name: str, timeout: int) → None[source]#: Join the NCCL weight sync group. Called by Synchronizer.

async teardown_weight_sync_group() → None[source]#: Destroy the NCCL weight sync group. Called by Synchronizer.

async train() → str[source]#: Train the model.

async train_step(exps: List[Experience]) → Dict[source]#

Train one step.

Returns:: Whether to continue training. Dict: Metrics of the training step.
Return type:: bool

async need_sync() → bool[source]#: Whether to sync the model weight.

need_save() → bool[source]#: Whether to save the checkpoint.

async sync_weight() → Dict[source]#: Sync the model weight.

async save_checkpoint(block_until_saved: bool = False, save_as_hf: bool = False) → Dict[source]#

async shutdown() → None[source]#

property train_step_num: int#: Get the current training step number.

async is_alive() → bool[source]#: Check if the trainer is alive.

classmethod get_actor(config: Config)[source]#: Get a Ray actor for the trainer.

class trinity.trainer.trainer.TrainEngineWrapper[source]#

Bases: ABC

A wrapper class to wrap various training engines.

abstractmethod async prepare() → None[source]#: Do some preparation before training started.

abstract property train_step_num: int#: Get the current training step number.

abstractmethod async train_step(batch_exps: List[Experience]) → Dict[source]#

Training one step.

Parameters:: batch_exps (List[Experience]) – A batch of experiences to train.
Returns:: Metrics of the training step.
Return type:: Dict

abstractmethod async save_checkpoint(block_until_saved: bool = False, save_as_hf: bool = False) → None[source]#: Save the whole checkpoint (Including model, optimizer, and other states).

async wait_for_save() → None[source]#

Wait for any pending background save operations to complete.

Default implementation is a no-op. Override in subclasses that use background save threads to ensure the checkpoint iteration file is written before the trainer exits.

abstractmethod sync_weight_nccl() → None[source]#: Sync the model weight by NCCL. (For NCCL sync method)

abstractmethod async upload_state_dict() → None[source]#: Upload the state dict to Synchronizer. (For MEMORY sync method)

abstractmethod async save_state_dict() → None[source]#: Only save the model state dict for Synchronizer. (For CHECKPOINT sync method)

abstractmethod async get_weight_sync_info() → Tuple[str, int, List] | None[source]#: Get (master_address, master_port, state_dict_meta) for NCCL group setup.

abstractmethod async setup_weight_sync_group(master_address: str, master_port: int, world_size: int, group_name: str, timeout: int) → None[source]#: Join the NCCL weight sync group.

abstractmethod async teardown_weight_sync_group() → None[source]#: Tear down the NCCL weight sync group.

trinity.trainer.trainer.is_verl_legacy() → bool[source]#: Return True when the installed verl package is < 0.8 (legacy backend).

trinity.trainer.trainer.get_latest_hf_checkpoint_path(config: Config) → str | None[source]#: Return the latest HF checkpoint path for a verl trainer config.

trinity.trainer.trainer.get_trainer_wrapper(config: Config) → TrainEngineWrapper[source]#: Get a trainer wrapper.

trinity.trainer.trainer module

Contents

trinity.trainer.trainer module#