Add support for DDP fork

docs/source-pytorch/accelerators/gpu_intermediate.rst

@@ -21,8 +21,10 @@ Lightning supports multiple ways of doing distributed training.
 |
 
 - Data Parallel (``strategy='dp'``) (multiple-gpus, 1 machine)
-- DistributedDataParallel (``strategy='ddp'``) (multiple-gpus across many machines (python script based)).
-- DistributedDataParallel (``strategy='ddp_spawn'``) (multiple-gpus across many machines (spawn based)).
+- DistributedDataParallel (multiple-gpus across many machines)
+    - Regular (``strategy='ddp'``)
+    - Spawn (``strategy='ddp_spawn'``)
+    - Fork (``strategy='ddp_fork'``)
 - Horovod (``strategy='horovod'``) (multi-machine, multi-gpu, configured at runtime)
 - Bagua (``strategy='bagua'``) (multiple-gpus across many machines with advanced training algorithms)
 
@@ -199,6 +201,61 @@ You can then call your scripts anywhere
     python some_file.py --accelerator 'gpu' --devices 8 --strategy 'ddp'
 
 
+Distributed Data Parallel Fork
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+DDP Fork is an alternative to Spawn that can be used in interactive Python and Jupyter notebooks, Google Colab, Kaggle notebooks, and so on:
+
+.. code-block:: python
+
+    # train on 8 GPUs in a Jupyter notebook
+    trainer = Trainer(accelerator="gpu", devices=8, strategy="ddp_fork")
+
+Data Parallel (``strategy="dp"``) is the only other strategy supported in interactive environments but is slower, is discouraged by PyTorch and has other limitations.
+Among the native distributed strategies, regular DDP (``strategy="ddp"``) is still recommended as the go-to strategy over Spawn and Fork for its speed and stability but it can only be used with scripts.
+
+
+Comparison of DDP variants and tradeoffs
+****************************************
+
+.. list-table:: DDP variants and their tradeoffs
+   :widths: 40 20 20 20
+   :header-rows: 1
+
+   * -
+     - DDP
+     - DDP Spawn
+     - DDP Fork
+   * - Works in Jupyter notebooks / IPython environments
+     - No
+     - No
+     - Yes
+   * - Supports multi-node
+     - Yes
+     - Yes
+     - Yes
+   * - Supported platforms
+     - Linux, Mac, Win
+     - Linux, Mac, Win
+     - Linux, Mac
+   * - Requires all objects to be picklable
+     - No
+     - Yes
+     - No
+   * - Is the guard ``if "__name__"=="__main__"`` required?
+     - Yes
+     - Yes
+     - No
+   * - Limitations in the main process
+     - None
+     - None
+     - GPU operations such as moving tensors to the GPU or calling ``torch.cuda`` functions before invoking ``Trainer.fit`` is not allowed.
+   * - Process creation time
+     - Slow
+     - Slow
+     - Fast
+
+
 Horovod
 ^^^^^^^
 `Horovod <http://horovod.ai>`_ allows the same training script to be used for single-GPU,

src/pytorch_lightning/CHANGELOG.md

@@ -99,6 +99,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Added Apple Silicon Support via `MPSAccelerator` ([#13123](https://github.com/PyTorchLightning/pytorch-lightning/pull/13123))
 
 
+- Added support for DDP Fork ([#13405](https://github.com/PyTorchLightning/pytorch-lightning/pull/13405))
+
 
 ### Changed
 

src/pytorch_lightning/accelerators/cuda.py

@@ -52,7 +52,7 @@ def setup(self, trainer: "pl.Trainer") -> None:
     def set_nvidia_flags(local_rank: int) -> None:
         # set the correct cuda visible devices (using pci order)
         os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
-        all_gpu_ids = ",".join(str(x) for x in range(torch.cuda.device_count()))
+        all_gpu_ids = ",".join(str(x) for x in range(device_parser.num_cuda_devices()))
         devices = os.getenv("CUDA_VISIBLE_DEVICES", all_gpu_ids)
         _log.info(f"LOCAL_RANK: {local_rank} - CUDA_VISIBLE_DEVICES: [{devices}]")
 
@@ -84,11 +84,11 @@ def get_parallel_devices(devices: List[int]) -> List[torch.device]:
     @staticmethod
     def auto_device_count() -> int:
         """Get the devices when set to auto."""
-        return torch.cuda.device_count()
+        return device_parser.num_cuda_devices()
 
     @staticmethod
     def is_available() -> bool:
-        return torch.cuda.device_count() > 0
+        return device_parser.num_cuda_devices() > 0
 
     @classmethod
     def register_accelerators(cls, accelerator_registry: Dict) -> None:
@@ -162,6 +162,6 @@ def _to_float(x: str) -> float:
 def _get_gpu_id(device_id: int) -> str:
     """Get the unmasked real GPU IDs."""
     # All devices if `CUDA_VISIBLE_DEVICES` unset
-    default = ",".join(str(i) for i in range(torch.cuda.device_count()))
+    default = ",".join(str(i) for i in range(device_parser.num_cuda_devices()))
     cuda_visible_devices = os.getenv("CUDA_VISIBLE_DEVICES", default=default).split(",")
     return cuda_visible_devices[device_id].strip()

src/pytorch_lightning/lite/lite.py

@@ -468,6 +468,7 @@ def _supported_strategy_types() -> Sequence[_StrategyType]:
             _StrategyType.DP,
             _StrategyType.DDP,
             _StrategyType.DDP_SPAWN,
+            _StrategyType.DDP_FORK,
             _StrategyType.DEEPSPEED,
             _StrategyType.DDP_SHARDED,
             _StrategyType.DDP_SHARDED_SPAWN,

src/pytorch_lightning/profilers/pytorch.py

@@ -24,6 +24,7 @@
 from torch.autograd.profiler import record_function
 
 from pytorch_lightning.profilers.profiler import Profiler
+from pytorch_lightning.utilities.device_parser import is_cuda_available
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.imports import _KINETO_AVAILABLE
 from pytorch_lightning.utilities.rank_zero import rank_zero_warn
@@ -368,7 +369,7 @@ def _default_activities(self) -> List["ProfilerActivity"]:
             return activities
         if self._profiler_kwargs.get("use_cpu", True):
             activities.append(ProfilerActivity.CPU)
-        if self._profiler_kwargs.get("use_cuda", torch.cuda.is_available()):
+        if self._profiler_kwargs.get("use_cuda", is_cuda_available()):
             activities.append(ProfilerActivity.CUDA)
         return activities
 

src/pytorch_lightning/strategies/ddp_spawn.py

@@ -22,6 +22,7 @@
 from torch.distributed.constants import default_pg_timeout
 from torch.nn import Module
 from torch.nn.parallel.distributed import DistributedDataParallel
+from typing_extensions import Literal
 
 import pytorch_lightning as pl
 from pytorch_lightning.overrides import LightningDistributedModule
@@ -71,6 +72,7 @@ def __init__(
         ddp_comm_wrapper: Optional[callable] = None,
         process_group_backend: Optional[str] = None,
         timeout: Optional[timedelta] = default_pg_timeout,
+        start_method: Literal["spawn", "fork", "forkserver"] = "spawn",
         **kwargs: Any,
     ):
         super().__init__(
@@ -88,6 +90,7 @@ def __init__(
         self._local_rank = 0
         self._process_group_backend: Optional[str] = process_group_backend
         self._timeout: Optional[timedelta] = timeout
+        self._start_method = start_method
 
     @property
     def num_nodes(self) -> int:
@@ -124,7 +127,7 @@ def process_group_backend(self) -> Optional[str]:
         return self._process_group_backend
 
     def _configure_launcher(self):
-        self._launcher = _SpawnLauncher(self)
+        self._launcher = _SpawnLauncher(self, start_method=self._start_method)
 
     def setup(self, trainer: "pl.Trainer") -> None:
         os.environ["MASTER_PORT"] = str(self.cluster_environment.main_port)
@@ -280,17 +283,20 @@ def post_training_step(self):
 
     @classmethod
     def register_strategies(cls, strategy_registry: Dict) -> None:
-        strategy_registry.register(
-            "ddp_spawn_find_unused_parameters_false",
-            cls,
-            description="DDPSpawn Strategy with `find_unused_parameters` as False",
-            find_unused_parameters=False,
-        )
-        strategy_registry.register(
-            cls.strategy_name,
-            cls,
-            description=f"{cls.__class__.__name__}",
-        )
+        for start_method in ("spawn", "fork"):
+            strategy_registry.register(
+                f"ddp_{start_method}_find_unused_parameters_false",
+                cls,
+                description=f"DDP {start_method.title()} strategy with `find_unused_parameters` as False",
+                find_unused_parameters=False,
+                start_method=start_method,
+            )
+            strategy_registry.register(
+                f"ddp_{start_method}",
+                cls,
+                description=f"DDP {start_method.title()} strategy",
+                start_method=start_method,
+            )
 
     def teardown(self) -> None:
         log.detail(f"{self.__class__.__name__}: tearing down strategy")

src/pytorch_lightning/strategies/launchers/spawn.py

@@ -20,6 +20,7 @@
 import torch
 import torch.multiprocessing as mp
 from torch import Tensor
+from typing_extensions import Literal
 
 import pytorch_lightning as pl
 from pytorch_lightning.strategies.launchers.base import _Launcher
@@ -34,27 +35,40 @@ class _SpawnLauncher(_Launcher):
     r"""Spawns processes that run a given function in parallel, and joins them all at the end.
 
     The main process in which this launcher is invoked creates N so-called worker processes (using
-    :func:`torch.multiprocessing.spawn`) that run the given function.
+    :func:`torch.multiprocessing.start_processes`) that run the given function.
     Worker processes have a rank that ranges from 0 to N - 1.
 
     Note:
         - This launcher requires all objects to be pickleable.
         - It is important that the entry point to the program/script is guarded by ``if __name__ == "__main__"``.
+        - With start method 'fork' the user must ensure that no CUDA context gets created in the main process before
+          the launcher is invoked. E.g., one should avoid creating cuda tensors or calling ``torch.cuda.*`` functions
+          before calling ``Trainer.fit``.
 
     Args:
         strategy: A reference to the strategy that is used together with this launcher.
+        start_method: The method how to start the processes.
+            - 'spawn': The default start method. Requires all objects to be pickleable.
+            - 'fork': Preferrable for IPython/Jupyter environments where 'spawn' is not available. Not available on
+              the Windows platform for example.
+            - 'forkserver': Alternative implementation to 'fork'.
     """
 
-    def __init__(self, strategy: Strategy) -> None:
+    def __init__(self, strategy: Strategy, start_method: Literal["spawn", "fork", "forkserver"] = "spawn") -> None:
         self._strategy = strategy
-        self._start_method = "spawn"
+        self._start_method = start_method
+        if start_method not in mp.get_all_start_methods():
+            raise ValueError(
+                f"The start method '{self._start_method}' is not available on this platform. Available methods are:"
+                f" {', '.join(mp.get_all_start_methods())}"
+            )
 
     @property
     def is_interactive_compatible(self) -> bool:
-        # The start method 'spawn' is currently the only one that works with DDP and CUDA support
-        # The start method 'fork' is the only one supported in Jupyter environments but not compatible with CUDA
-        # For more context, see https://github.com/Lightning-AI/lightning/issues/7550
-        return self._start_method == "fork" and self._strategy.root_device.type != "cuda"
+        # The start method 'spawn' is not supported in interactive environments
+        # The start method 'fork' is the only one supported in Jupyter environments, with constraints around CUDA
+        # initialization. For more context, see https://github.com/Lightning-AI/lightning/issues/7550
+        return self._start_method == "fork"
 
     def launch(self, function: Callable, *args: Any, trainer: Optional["pl.Trainer"] = None, **kwargs: Any) -> Any:
         """Spawns processes that run the given function in parallel.
@@ -75,7 +89,7 @@ def launch(self, function: Callable, *args: Any, trainer: Optional["pl.Trainer"]
         os.environ["MASTER_PORT"] = str(self._strategy.cluster_environment.main_port)
         context = mp.get_context(self._start_method)
         return_queue = context.SimpleQueue()
-        mp.spawn(
+        mp.start_processes(
             self._wrapping_function,
             args=(trainer, function, args, kwargs, return_queue),
             nprocs=self._strategy.num_processes,

src/pytorch_lightning/strategies/launchers/xla_spawn.py

@@ -50,8 +50,7 @@ class _XLASpawnLauncher(_SpawnLauncher):
     """
 
     def __init__(self, strategy: "Strategy") -> None:
-        super().__init__(strategy)
-        self._start_method = "fork"
+        super().__init__(strategy=strategy, start_method="fork")
 
     @property
     def is_interactive_compatible(self) -> bool:

src/pytorch_lightning/strategies/tpu_spawn.py

@@ -70,9 +70,9 @@ def __init__(
             cluster_environment=XLAEnvironment(),
             checkpoint_io=checkpoint_io,
             precision_plugin=precision_plugin,
+            start_method="fork",
         )
         self.debug = debug
-        self.start_method = "fork"
 
     @property
     def root_device(self) -> torch.device:
@@ -113,7 +113,6 @@ def _configure_launcher(self):
         self._launcher = _XLASpawnLauncher(self)
 
     def setup(self, trainer: "pl.Trainer") -> None:
-        self.start_method = "fork"
         self.accelerator.setup(trainer)
 
         if self.debug:

src/pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -267,7 +267,7 @@ def _check_config_and_set_final_flags(
             if strategy == "ddp_cpu":
                 raise MisconfigurationException(
                     "`Trainer(strategy='ddp_cpu')` is not a valid strategy,"
-                    " you can use `Trainer(strategy='ddp'|'ddp_spawn', accelerator='cpu')` instead."
+                    " you can use `Trainer(strategy='ddp'|'ddp_spawn'|'ddp_fork', accelerator='cpu')` instead."
                 )
             if strategy == "tpu_spawn":
                 raise MisconfigurationException(
@@ -496,7 +496,7 @@ def _choose_accelerator(self) -> str:
                 return "hpu"
             if MPSAccelerator.is_available():
                 return "mps"
-            if torch.cuda.is_available() and torch.cuda.device_count() > 0:
+            if CUDAAccelerator.is_available():
                 return "cuda"
         return "cpu"
 
@@ -614,7 +614,14 @@ def _check_strategy_and_fallback(self) -> None:
                 f"You selected strategy to be `{DDPFullyShardedNativeStrategy.strategy_name}`, "
                 "but GPU accelerator is not used."
             )
-
+        if (
+            strategy_flag in ("ddp_fork", "ddp_fork_find_unused_parameters_false")
+            and "fork" not in torch.multiprocessing.get_all_start_methods()
+        ):
+            raise ValueError(
+                f"You selected `Trainer(strategy='{strategy_flag}')` but process forking is not supported on this"
+                f" platform. We recommed `Trainer(strategy='ddp_spawn')` instead."
+            )
         if strategy_flag:
             self._strategy_flag = strategy_flag
 

src/pytorch_lightning/trainer/trainer.py

@@ -1758,7 +1758,7 @@ def _log_device_info(self) -> None:
         rank_zero_info(f"HPU available: {_HPU_AVAILABLE}, using: {num_hpus} HPUs")
 
         # TODO: Integrate MPS Accelerator here, once gpu maps to both
-        if torch.cuda.is_available() and not isinstance(self.accelerator, CUDAAccelerator):
+        if CUDAAccelerator.is_available() and not isinstance(self.accelerator, CUDAAccelerator):
             rank_zero_warn(
                 "GPU available but not used. Set `accelerator` and `devices` using"
                 f" `Trainer(accelerator='gpu', devices={CUDAAccelerator.auto_device_count()})`.",

src/pytorch_lightning/tuner/auto_gpu_select.py

@@ -15,6 +15,7 @@
 
 import torch
 
+from pytorch_lightning.utilities import device_parser
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
 
 
@@ -31,7 +32,7 @@ def pick_multiple_gpus(nb: int) -> List[int]:
             " Please select a valid number of GPU resources when using auto_select_gpus."
         )
 
-    num_gpus = torch.cuda.device_count()
+    num_gpus = device_parser.num_cuda_devices()
     if nb > num_gpus:
         raise MisconfigurationException(f"You requested {nb} GPUs but your machine only has {num_gpus} GPUs.")
     nb = num_gpus if nb == -1 else nb
@@ -51,7 +52,7 @@ def pick_single_gpu(exclude_gpus: List[int]) -> int:
     """
     previously_used_gpus = []
     unused_gpus = []
-    for i in range(torch.cuda.device_count()):
+    for i in range(device_parser.num_cuda_devices()):
         if i in exclude_gpus:
             continue