verl/trainer/fsdp_sft_trainer.py

• extract_step · function · L52-L56 — def extract_step(path)
• convert_to_regular_types · function · L59-L68 — def convert_to_regular_types(obj)
• FSDPSFTTrainer · class · L71-L414 — class FSDPSFTTrainer(object)
• init · method · L73-L92 — def init(self, config, device_mesh: DeviceMesh)
• _normalize_config_bsz · method · L94-L103 — def _normalize_config_bsz(self)
• _build_dataloader · method · L105-L150 — def _build_dataloader(self)
• _build_model_optimizer · method · L152-L242 — def _build_model_optimizer(self): # TODO (zhangchi.usc1992): # 1. support pretrain from random weights # 2. support init directly from sharded weights
• _compute_loss · method · L244-L276 — def _compute_loss(self, batch)
• training_step · method · L278-L312 — def training_step(self, batch: TensorDict)
• validation_step · method · L314-L319 — def validation_step(self, batch: TensorDict)
• save_checkpoint · method · L321-L337 — def save_checkpoint(self, step): # save checkpoint
• fit · method · L339-L414 — def fit(self)
• main · function · L426-L431 — def main(config)