[tune] distributed torch wrapper (#9550)

* changes

* add-working

* checkpoint

* ccleanu

* fix

* ok

* formatting

* ok

* tests

* some-good-stuff

* fix-torch

* ddp-torch

* torch-test

* sessions

* add-small-test

* fix

* remove

* gpu-working

* update-tests

* ok

* try-test

* formgat

* ok

* ok

python/ray/util/sgd/BUILD

@@ -26,6 +26,14 @@ py_test(
     deps = [":sgd_lib"],
 )
 
+py_test(
+    name = "test_torch_trainable",
+    size = "small",
+    srcs = ["tests/test_torch_trainable.py"],
+    tags = ["exclusive", "pytorch"],
+    deps = [":sgd_lib"],
+)
+
 # --------------------------------------------------------------------
 # Tests from the python/ray/util/sgd/tf/examples directory.
 # Please keep these sorted alphabetically.
@@ -181,6 +189,7 @@ py_test(
     args = ["--num-workers=2"]
 )
 
+
 # --------------------------------------------------------------------
 # Tests from the python/ray/util/sgd/torch/examples/* directories.
 # Only covers subdirectories.

python/ray/util/sgd/tests/test_torch_trainable.py

@@ -0,0 +1,85 @@
+import os
+import pytest
+from unittest.mock import patch
+import torch
+import torch.distributed as dist
+
+import ray
+from ray import tune
+from ray.util.sgd.torch.func_trainable import (
+    DistributedTrainableCreator, distributed_checkpoint, _train_simple)
+
+
+@pytest.fixture
+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()
+
+
+@pytest.fixture
+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()
+
+
+def test_single_step(ray_start_2_cpus):  # noqa: F811
+    trainable_cls = DistributedTrainableCreator(_train_simple, num_workers=2)
+    trainer = trainable_cls()
+    trainer.train()
+    trainer.stop()
+
+
+def test_step_after_completion(ray_start_2_cpus):  # noqa: F811
+    trainable_cls = DistributedTrainableCreator(_train_simple, num_workers=2)
+    trainer = trainable_cls(config={"epochs": 1})
+    with pytest.raises(RuntimeError):
+        for i in range(10):
+            trainer.train()
+
+
+def test_save_checkpoint(ray_start_2_cpus):  # noqa: F811
+    trainable_cls = DistributedTrainableCreator(_train_simple, num_workers=2)
+    trainer = trainable_cls(config={"epochs": 1})
+    trainer.train()
+    path = trainer.save()
+    model_state_dict, opt_state_dict = torch.load(path)
+    trainer.stop()
+
+
+@pytest.mark.parametrize("enabled_checkpoint", [True, False])
+def test_simple_tune(ray_start_4_cpus, enabled_checkpoint):
+    trainable_cls = DistributedTrainableCreator(_train_simple, num_workers=2)
+    analysis = tune.run(
+        trainable_cls,
+        config={"enable_checkpoint": enabled_checkpoint},
+        num_samples=2,
+        stop={"training_iteration": 2})
+    assert analysis.trials[0].last_result["training_iteration"] == 2
+    assert analysis.trials[0].has_checkpoint() == enabled_checkpoint
+
+
+@pytest.mark.parametrize("rank", [0, 1])
+def test_checkpoint(ray_start_2_cpus, rank):  # noqa: F811
+    with patch("torch.distributed.get_rank") as rank_method:
+        rank_method.return_value = rank
+        with distributed_checkpoint(label="test") as path:
+            if rank == 0:
+                assert path
+            else:
+                assert path == os.devnull
+
+
+if __name__ == "__main__":
+    import pytest
+    import sys
+    sys.exit(pytest.main(["-v", __file__]))

python/ray/util/sgd/torch/__init__.py

@@ -12,7 +12,13 @@ try:
                                                   BaseTorchTrainable)
 
     from ray.util.sgd.torch.training_operator import TrainingOperator
+    from ray.util.sgd.torch.func_trainable import (DistributedTrainableCreator,
+                                                   distributed_checkpoint)
 
-    __all__ = ["TorchTrainer", "BaseTorchTrainable", "TrainingOperator"]
-except ImportError:
+    __all__ = [
+        "TorchTrainer", "BaseTorchTrainable", "TrainingOperator",
+        "distributed_checkpoint", "DistributedTrainableCreator"
+    ]
+except ImportError as e:
+    logger.warning(e)
     logger.warning("PyTorch not found. TorchTrainer will not be available")

python/ray/util/sgd/torch/distributed_torch_runner.py

@@ -1,4 +1,3 @@
-from datetime import timedelta
 import logging
 import io
 import os
@@ -8,7 +7,7 @@ import torch.distributed as dist
 from torch.nn.parallel import DistributedDataParallel
 from torch.utils.data import DataLoader, IterableDataset
 from torch.utils.data.distributed import DistributedSampler
-from ray.util.sgd.torch.constants import NCCL_TIMEOUT_S
+from ray.util.sgd.torch.utils import setup_process_group
 
 import ray
 from ray.util.sgd.torch.torch_runner import TorchRunner
@@ -47,31 +46,19 @@ class DistributedTorchRunner(TorchRunner):
     def setup(self):
         raise RuntimeError("Need to call setup commands separately.")
 
-    def setup_process_group(self, url, world_rank, world_size):
+    def setup_process_group(self, url, world_rank, world_size, timeout):
         """Connects the distributed PyTorch backend.
 
         Args:
             url (str): the URL used to connect to distributed PyTorch.
             world_rank (int): the index of the runner.
             world_size (int): the total number of runners.
+            timeout (timedelta): Seconds for process group
+                operations to timeout.
         """
         self.world_rank = world_rank
-        logger.debug("Connecting to {} world_rank: {} world_size: {}".format(
-            url, world_rank, world_size))
-        logger.debug("using {}".format(self.backend))
-        if self.backend == "nccl" and "NCCL_BLOCKING_WAIT" not in os.environ:
-            logger.debug(
-                "Setting NCCL_BLOCKING_WAIT for detecting node failure. "
-                "To override this behavior, you can set NCCL_BLOCKING_WAIT=0.")
-            os.environ["NCCL_BLOCKING_WAIT"] = "1"
-
-        timeout = timedelta(seconds=NCCL_TIMEOUT_S)
-        dist.init_process_group(
-            backend=self.backend,
-            init_method=url,
-            rank=world_rank,
-            world_size=world_size,
-            timeout=timeout)
+        setup_process_group(
+            url, world_rank, world_size, timeout, backend=self.backend)
 
     def setup_ddp_and_operator(self):
         """Runs distributed coordination components.

python/ray/util/sgd/torch/func_trainable.py

@@ -0,0 +1,235 @@
+# Original Code here:
+# https://github.com/pytorch/examples/blob/master/mnist/main.py
+import os
+import logging
+import torch
+from datetime import timedelta
+
+import ray
+from ray import tune
+from ray.tune.result import RESULT_DUPLICATE
+from ray.tune.logger import NoopLogger
+from ray.tune.function_runner import wrap_function
+from ray.tune.resources import Resources
+from ray.tune.trainable import TrainableUtil
+from ray.util.sgd.torch.utils import setup_process_group
+from ray.util.sgd.torch.constants import NCCL_TIMEOUT_S
+from ray.util.sgd.torch.utils import setup_address
+
+logger = logging.getLogger(__name__)
+
+
+def logger_creator(log_config, logdir, rank):
+    worker_dir = os.path.join(logdir, "worker_{}".format(rank))
+    os.makedirs(worker_dir, exist_ok=True)
+    return NoopLogger(log_config, worker_dir)
+
+
+class _TorchTrainable(tune.Trainable):
+    """Base class for distributed training on Tune.
+
+    A wrapper class is needed to actually create a working
+    version of this trainable.
+    """
+    _function = None
+    _num_workers = None
+    _use_gpu = None
+    _num_cpus_per_worker = None
+
+    __slots__ = ["workers", "_finished"]
+
+    @classmethod
+    def default_process_group_parameters(self):
+        return dict(timeout=timedelta(NCCL_TIMEOUT_S), backend="gloo")
+
+    @classmethod
+    def get_remote_worker_options(self):
+        num_gpus = 1 if self._use_gpu else 0
+        num_cpus = int(self._num_cpus_per_worker or 1)
+        return dict(num_cpus=num_cpus, num_gpus=num_gpus)
+
+    def setup(self, config):
+        self._finished = False
+        num_workers = self._num_workers
+        logdir = self.logdir
+        assert self._function
+
+        func_trainable = wrap_function(self.__class__._function)
+
+        remote_trainable = ray.remote(func_trainable)
+        remote_trainable = remote_trainable.options(
+            **self.get_remote_worker_options())
+
+        address = setup_address()
+        self.workers = [
+            remote_trainable.remote(
+                config=config,
+                logger_creator=lambda cfg: logger_creator(cfg, logdir, rank))
+            for rank in range(num_workers)
+        ]
+
+        pgroup_params = self.default_process_group_parameters()
+        from functools import partial
+        setup_on_worker = partial(
+            setup_process_group,
+            url=address,
+            world_size=num_workers,
+            **pgroup_params)
+        ray.get([
+            w.execute.remote(lambda _: setup_on_worker(world_rank=rank))
+            for rank, w in enumerate(self.workers)
+        ])
+
+    def step(self):
+        if self._finished:
+            raise RuntimeError("Training has already finished.")
+        result = ray.get([w.step.remote() for w in self.workers])[0]
+        if RESULT_DUPLICATE in result:
+            self._finished = True
+        return result
+
+    def save_checkpoint(self, checkpoint_dir):
+        # TODO: optimize if colocated
+        save_obj = ray.get(self.workers[0].save_to_object.remote())
+        checkpoint_path = TrainableUtil.create_from_pickle(
+            save_obj, checkpoint_dir)
+        return checkpoint_path
+
+    def load_checkpoint(self, checkpoint_dir):
+        checkpoint_obj = TrainableUtil.checkpoint_to_object(checkpoint_dir)
+        return ray.get(
+            w.restore_from_object.remote(checkpoint_obj) for w in self.workers)
+
+    def stop(self):
+        ray.get([worker.stop.remote() for worker in self.workers])
+
+
+def DistributedTrainableCreator(func,
+                                use_gpu=False,
+                                num_workers=1,
+                                num_cpus_per_worker=1,
+                                backend="gloo",
+                                timeout_s=NCCL_TIMEOUT_S):
+    """Creates a class that executes distributed training.
+
+    Note that you typically should not instantiate the object
+    created.
+
+    Example:
+
+    .. code-block::
+
+        trainable_cls = DistributedTrainableCreator(
+            train_func, num_workers=2)
+        analysis = tune.run(trainable_cls)
+    """
+
+    class WrappedDistributedTorchTrainable(_TorchTrainable):
+        _function = func
+        _num_workers = num_workers
+        _use_gpu = use_gpu
+        _num_cpus_per_worker = num_cpus_per_worker
+
+        @classmethod
+        def default_process_group_parameters(self):
+            return dict(timeout=timedelta(timeout_s), backend=backend)
+
+        @classmethod
+        def default_resource_request(cls, config):
+            num_workers_ = int(config.get("num_workers", num_workers))
+            num_cpus = int(
+                config.get("num_cpus_per_worker", num_cpus_per_worker))
+            use_gpu_ = config.get("use_gpu", use_gpu)
+
+            return Resources(
+                cpu=0,
+                gpu=0,
+                extra_cpu=num_cpus * num_workers_,
+                extra_gpu=num_workers_ if use_gpu_ else 0)
+
+    return WrappedDistributedTorchTrainable
+
+
+class distributed_checkpoint:
+    """ContextManager for creating a distributed checkpoint.
+
+    Only checkpoints a file on the "main" training actor, avoiding
+    redundant work.
+
+    Args:
+        label (int | str): Used to label the checkpoint
+        disable (bool): Disable for prototyping.
+
+    Example:
+
+    .. code-block::
+
+        if epoch % 3 == 0:
+            with distributed_checkpoint(label=epoch) as path:
+                torch.save(model.state_dict(), path)
+    """
+
+    def __init__(self, label, disable=False):
+        self.label = label
+        self.file = None
+        self.disable = disable
+
+    def __enter__(self):
+        if torch.distributed.get_rank() == 0 and not self.disable:
+            checkpoint_dir = tune.make_checkpoint_dir(step=self.label)
+            path = os.path.join(checkpoint_dir, "checkpoint")
+        else:
+            path = os.devnull
+        self.file = path
+        return path
+
+    def __exit__(self, type, value, traceback):
+        if torch.distributed.get_rank() == 0 and not self.disable:
+            tune.save_checkpoint(self.file)
+
+
+def _train_simple(config, checkpoint=False):
+    """For testing only. Putting this here because Ray has problems
+    serializing within the test file."""
+    import torch.nn as nn
+    from torch.nn.parallel import DistributedDataParallel
+    import torch.optim as optim
+    # N is batch size; D_in is input dimension;
+    # H is hidden dimension; D_out is output dimension.
+    N, D_in, H, D_out = 8, 5, 5, 5
+
+    # Create random Tensors to hold inputs and outputs
+    x = torch.randn(N, D_in)
+    y = torch.randn(N, D_out)
+    loss_fn = nn.MSELoss()
+
+    # Use the nn package to define our model and loss function.
+    model = torch.nn.Sequential(
+        torch.nn.Linear(D_in, H),
+        torch.nn.ReLU(),
+        torch.nn.Linear(H, D_out),
+    )
+    optimizer = optim.SGD(model.parameters(), lr=0.1)
+
+    if checkpoint:
+        with open(checkpoint) as f:
+            model_state, optimizer_state = torch.load(f)
+
+        model.load_state_dict(model_state)
+        optimizer.load_state_dict(optimizer_state)
+
+    model = DistributedDataParallel(model)
+
+    for epoch in range(config.get("epochs", 10)):
+        optimizer.zero_grad()
+        output = model(x)
+        loss = loss_fn(output, y)
+        loss.backward()
+        optimizer.step()
+
+        if epoch % 3 == 0:
+            if config.get("enable_checkpoint", True):
+                with distributed_checkpoint(label=epoch) as path:
+                    torch.save((model.state_dict(), optimizer.state_dict()),
+                               path)
+        tune.report(mean_loss=loss.item())

python/ray/util/sgd/torch/torch_trainer.py

@@ -1,3 +1,4 @@
+from datetime import timedelta
 import numpy as np
 import logging
 import os
@@ -16,7 +17,8 @@ from ray.util.sgd.torch.distributed_torch_runner import (
     DistributedTorchRunner, LocalDistributedRunner, DeactivatedRunner)
 from ray.util.sgd.utils import check_for_failure, NUM_SAMPLES, BATCH_SIZE
 from ray.util.sgd.torch.torch_runner import TorchRunner
-from ray.util.sgd.torch.constants import VALID_SCHEDULER_STEP
+from ray.util.sgd.torch.constants import VALID_SCHEDULER_STEP, NCCL_TIMEOUT_S
+from ray.util.sgd.torch.utils import setup_address
 from ray.util.sgd.data import Dataset
 
 logger = logging.getLogger(__name__)
@@ -180,6 +182,7 @@ class TorchTrainer:
             use_gpu="auto",
             backend="auto",
             wrap_ddp=True,
+            timeout_s=NCCL_TIMEOUT_S,
             serialize_data_creation=True,
             use_fp16=False,
             use_tqdm=False,
@@ -248,6 +251,7 @@ class TorchTrainer:
 
         self.serialize_data_creation = serialize_data_creation
         self.wrap_ddp = wrap_ddp
+        self.timeout_s = timeout_s
         self.use_fp16 = use_fp16
         self.use_tqdm = use_tqdm
         self.add_dist_sampler = add_dist_sampler
@@ -347,10 +351,7 @@ class TorchTrainer:
                 self.apply_all_workers(self.initialization_hook)
 
             # Compute URL for initializing distributed PyTorch
-            ip = ray.services.get_node_ip_address()
-            port = self.local_worker.find_free_port()
-
-            address = "tcp://{ip}:{port}".format(ip=ip, port=port)
+            address = setup_address()
 
             # Runs the creator functions.
             remote_component_setup = [
@@ -363,10 +364,12 @@ class TorchTrainer:
 
             # Setup the process group among all workers.
             remote_pgroup_setups = [
-                worker.setup_process_group.remote(address, i + 1, num_workers)
+                worker.setup_process_group.remote(address, i + 1, num_workers,
+                                                  timedelta(self.timeout_s))
                 for i, worker in enumerate(self.remote_workers)
             ]
-            self.local_worker.setup_process_group(address, 0, num_workers)
+            self.local_worker.setup_process_group(address, 0, num_workers,
+                                                  timedelta(self.timeout_s))
             # Get setup tasks in order to throw errors on failure
             ray.get(remote_pgroup_setups)
 

python/ray/util/sgd/torch/utils.py

@@ -0,0 +1,43 @@
+import os
+import logging
+import torch.distributed as dist
+
+import ray
+from ray.util.sgd.utils import find_free_port
+
+logger = logging.getLogger(__name__)
+
+
+def setup_address():
+    ip = ray.services.get_node_ip_address()
+    port = find_free_port()
+    return "tcp://{ip}:{port}".format(ip=ip, port=port)
+
+
+def setup_process_group(url, world_rank, world_size, timeout, backend="gloo"):
+    """Connects the distributed PyTorch backend.
+
+    Args:
+        url (str): the URL used to connect to distributed PyTorch.
+        world_rank (int): the index of the runner.
+        world_size (int): the total number of runners.
+        timeout (timedelta): Timeout for operations executed against
+            the process group.
+        backend (str): One of gloo or nccl. Depending on
+            build-time configuration
+    """
+    logger.debug("Connecting to {} world_rank: {} world_size: {}".format(
+        url, world_rank, world_size))
+    logger.debug("using {}".format(backend))
+    if backend == "nccl" and "NCCL_BLOCKING_WAIT" not in os.environ:
+        logger.debug(
+            "Setting NCCL_BLOCKING_WAIT for detecting node failure. "
+            "To override this behavior, you can set NCCL_BLOCKING_WAIT=0.")
+        os.environ["NCCL_BLOCKING_WAIT"] = "1"
+
+    dist.init_process_group(
+        backend=backend,
+        init_method=url,
+        rank=world_rank,
+        world_size=world_size,
+        timeout=timeout)