Method init_process_group

checkpoint_engine/ps.py:502–523

Initialize the process group for the ranks. This global group can be easily destroyed by calling dist.destroy_process_group. Args: master_port: The specified port of the master node. If not set, will use _get_master_port to get the port. timeout: The timeout

(
        self,
        *,
        timeout: timedelta = timedelta(minutes=10),
    )

        )

    def init_process_group(
        self,
        *,
        timeout: timedelta = timedelta(minutes=10),
    ):
        """
        Initialize the process group for the ranks. This global group can be easily destroyed by calling dist.destroy_process_group.

        Args:
            master_port: The specified port of the master node. If not set, will use _get_master_port to get the port.
            timeout: The timeout of the process group.
        """
        self._store_counter += 1
        sub_store = torch.distributed.PrefixStore(f"prefix-{self._store_counter}", self._store)
        dist.init_process_group(
            backend=self.device_manager.backend,
            world_size=self._world_size,
            rank=self._rank,
            timeout=timeout,
            store=sub_store,
        )
        logger.info(f"[rank{self._rank}] init process group successfully.")

    def store_based_barrier(self, timeout: timedelta = timedelta(minutes=5)) -> None:
        """