← Về thư mục
verl/workers/megatron_workers.py
- set_random_seed · function · L47-L60 — def set_random_seed(seed)
- ActorRolloutRefWorker · class · L63-L413 — class ActorRolloutRefWorker(MegatronWorker)
- init · method · L69-L122 — def init(self, config: DictConfig, role: str)
- _build_model_optimizer · method · L124-L214 — def _build_model_optimizer(self, model_path, megatron_config: ModelParallelConfig, optim_config, override_model_config, enable_gradient_checkpointing=False)
- megatron_actor_model_provider · function · L154-L166 — def megatron_actor_model_provider(pre_process, post_process)
- _build_rollout · method · L216-L258 — def _build_rollout(self)
- init_model · method · L261-L323 — def init_model(self)
- update_actor · method · L326-L342 — def update_actor(self, data: DataProto)
- generate_sequences · method · L353-L381 — def generate_sequences(self, prompts: DataProto)
- compute_ref_log_prob · method · L384-L400 — def compute_ref_log_prob(self, data: DataProto)
- load_checkpoint · method · L403-L404 — def load_checkpoint(self, checkpoint_path)
- load_pretrained_model · method · L407-L408 — def load_pretrained_model(self, checkpoint_path)
- save_checkpoint · method · L411-L413 — def save_checkpoint(self, checkpoint_path)
- CriticWorker · class · L416-L581 — class CriticWorker(MegatronWorker)
- init · method · L418-L452 — def init(self, config)
- _build_critic_model_optimizer · method · L454-L519 — def _build_critic_model_optimizer(self, model_path, megatron_config: ModelParallelConfig, optim_config, override_model_config, enable_gradient_checkpointing=False)
- megatron_critic_model_provider · function · L484-L496 — def megatron_critic_model_provider(pre_process, post_process)
- init_model · method · L522-L556 — def init_model(self): # create critic
- compute_values · method · L559-L564 — def compute_values(self, data: DataProto)
- update_critic · method · L567-L573 — def update_critic(self, data: DataProto)
- load_checkpoint · method · L576-L577 — def load_checkpoint(self, checkpoint_path)
- save_checkpoint · method · L580-L581 — def save_checkpoint(self, checkpoint_path)
- RewardModelWorker · class · L584-L735 — class RewardModelWorker(MegatronWorker)
- init · method · L589-L620 — def init(self, config)
- _build_rm_model · method · L622-L677 — def _build_rm_model(self, model_path, megatron_config: ModelParallelConfig, override_model_config)
- megatron_rm_model_provider · function · L646-L658 — def megatron_rm_model_provider(pre_process, post_process)
- init_model · method · L680-L726 — def init_model(self): # create critic
- compute_rm_score · method · L731-L735 — def compute_rm_score(self, data: DataProto)