Python distributed.destroy_process_group方法代码示例

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def spmd_main(local_world_size, local_rank):
    # These are the parameters used to initialize the process group
    env_dict = {
        key: os.environ[key]
        for key in ("MASTER_ADDR", "MASTER_PORT", "RANK", "WORLD_SIZE")
    }
    print(f"[{os.getpid()}] Initializing process group with: {env_dict}")
    dist.init_process_group(backend="nccl")
    print(
        f"[{os.getpid()}]: world_size = {dist.get_world_size()}, "
        + f"rank = {dist.get_rank()}, backend={dist.get_backend()}"
    )

    demo_basic(local_world_size, local_rank)

    # Tear down the process group
    dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def _test__native_dist_model_create_from_context_no_dist(true_backend, true_device):

    assert _NativeDistModel.create_from_context() is None

    dist.init_process_group(true_backend, "tcp://0.0.0.0:2222", world_size=1, rank=0)
    dist.barrier()

    _test__native_dist_model_create_from_context_no_local_rank()

    true_conf = {
        "device": true_device,
        "local_rank": 0,
        "rank": 0,
        "world_size": 1,
        "node_index": 0,
        "nnodes": 1,
        "nproc_per_node": 1,
    }

    _test__native_dist_model_create_from_context_env_local_rank(true_conf)
    _test__native_dist_model_create_from_context_set_local_rank(true_conf)

    dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def clean_mp(self):
        if self.can_parallel:
            dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def cleanup():
    dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def close(self):
        r"""关闭Trainer，销毁进程"""
        dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def cleanup():
    dist.destroy_process_group()

# pylint: disable=no-member 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def shutdown(cls):
        if dist.get_rank() == 0 and cls.instance.ttp_initialized:
            cls.instance.send_obj(
                "terminate", cls.instance.get_ttp_rank(), cls.instance.ttp_group
            )
        dist.destroy_process_group(cls.instance.main_group)
        dist.destroy_process_group(cls.instance.ttp_group)
        dist.destroy_process_group()
        cls.instance = None 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def tearDown(self):
        # we need to de-initialize the distributed world, otherwise other
        # tests will they're we're distributed when we're really not.
        dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def ray_start_2_cpus():
    address_info = ray.init(num_cpus=2)
    yield address_info
    # The code after the yield will run as teardown code.
    ray.shutdown()
    # Ensure that tests don't ALL fail
    if dist.is_initialized():
        dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def ray_start_4_cpus():
    address_info = ray.init(num_cpus=4)
    yield address_info
    # The code after the yield will run as teardown code.
    ray.shutdown()
    # Ensure that tests don't ALL fail
    if dist.is_initialized():
        dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def shutdown(self):
        """Attempts to shut down the worker."""
        # However, it seems to be harmless to remove permanently
        # since the processes are shutdown anyways. This comment can be
        # removed in a future release if it is still not documented
        # the stable Pytorch docs.
        dist.destroy_process_group()
        super(DistributedTorchRunner, self).shutdown() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def run_training_teardown(self):
        if hasattr(self, '_teardown_already_run') and self._teardown_already_run:
            return

        self._teardown_already_run = True

        # Train end events
        with self.profiler.profile('on_train_end'):
            # callbacks
            self.on_train_end()
            # model hooks
            if self.is_function_implemented('on_train_end'):
                self.get_model().on_train_end()

        if self.logger is not None:
            self.logger.finalize("success")

        # summarize profile results
        if self.global_rank == 0:
            self.profiler.describe()

        if self.global_rank == 0:
            for proc in self.interactive_ddp_procs:
                subprocess.Popen.kill(proc)

        # clean up dist group
        if self.use_ddp or self.use_ddp2:
            torch_distrib.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def test_sync_reduce_ddp():
    """Make sure sync-reduce works with DDP"""
    tutils.reset_seed()
    tutils.set_random_master_port()

    worldsize = 2
    mp.spawn(_ddp_test_fn, args=(worldsize,), nprocs=worldsize)

    # dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def test_tensor_metric_ddp():
    tutils.reset_seed()
    tutils.set_random_master_port()

    world_size = 2
    mp.spawn(_ddp_test_tensor_metric, args=(world_size,), nprocs=world_size)
    # dist.destroy_process_group() 

# 需要导入模块: from torch import distributed [as 别名]
# 或者: from torch.distributed import destroy_process_group [as 别名]
def test_numpy_metric_ddp():
    tutils.reset_seed()
    tutils.set_random_master_port()
    world_size = 2
    mp.spawn(_ddp_test_numpy_metric, args=(world_size,), nprocs=world_size)
    # dist.destroy_process_group()