← Về thư mục
verl/workers/critic/megatron_critic.py
- MegatronPPOCritic · class · L41-L229 — class MegatronPPOCritic(BasePPOCritic)
- init · method · L43-L75 — def init(self, config, model_config, megatron_config, critic_module: nn.ModuleList, critic_optimizer: DistributedOptimizer, critic_optimizer_config: OptimizerConfig)
- compute_values · method · L77-L104 — def compute_values(self, data: DataProto) -> DataProto: # data.batch = data.batch.to(self.critic_module.module.device)
- make_minibatch_iterator · method · L106-L111 — def make_minibatch_iterator(self, data: DataProto) -> Iterable[DataProto]
- forward_backward_batch · method · L113-L202 — def forward_backward_batch(self, data: DataProto, forward_only=False): # broadcast from last pp rank to all other pp ranks
- loss_func · function · L135-L163 — def loss_func(output, data, meta_info)
- forward_step · function · L165-L171 — def forward_step(batch_iter, model)
- update_critic · method · L204-L229 — def update_critic(self, dataloader: Iterable[DataProto])