Fix sync_dist for tpus (#6950)

(cherry picked from commit 1b3e4f9)

CHANGELOG.md

@@ -237,6 +237,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed `AttributeError for `require_backward_grad_sync` when running manual optimization with sharded plugin ([#6915](https://github.com/PyTorchLightning/pytorch-lightning/pull/6915))
 
 
+- Fixed `sync_dist` for tpus ([#6950](https://github.com/PyTorchLightning/pytorch-lightning/pull/6950))
+
+
 ## [1.2.7] - 2021-04-06
 
 ### Fixed

pytorch_lightning/core/step_result.py

@@ -22,7 +22,7 @@
 from torch import Tensor
 from torchmetrics import Metric
 
-from pytorch_lightning.utilities.distributed import sync_ddp_if_available
+from pytorch_lightning.utilities.distributed import sync_ddp_if_available, tpu_distributed
 
 
 class Result(Dict):
@@ -139,10 +139,11 @@ def log(
 
         # sync across workers when using distributed training
         sync_fn = sync_fn or sync_ddp_if_available
+
         if sync_dist and isinstance(value, (torch.Tensor, numbers.Number)):
             is_dist_initialized = torch.distributed.is_available() and torch.distributed.is_initialized()
             # TODO: Find a way to make the reduction only once, so we don't need to clone.
-            if is_dist_initialized and isinstance(value, torch.Tensor):
+            if (is_dist_initialized or tpu_distributed) and isinstance(value, torch.Tensor):
                 value = value.clone()
             else:
                 value = torch.tensor(value, device=device, dtype=torch.float)

pytorch_lightning/plugins/training_type/tpu_spawn.py

@@ -15,7 +15,7 @@
 import os
 import re
 import time
-from typing import Any, Dict, List, Optional, Union, TYPE_CHECKING
+from typing import Any, Dict, List, Optional, TYPE_CHECKING, Union
 
 import torch
 import torch.multiprocessing as mp
@@ -40,7 +40,6 @@
 if _OMEGACONF_AVAILABLE:
     from omegaconf import DictConfig, ListConfig, OmegaConf
 
-
 if TYPE_CHECKING:
     from torch.nn import Module
     from torch.utils.data import DataLoader
@@ -276,4 +275,6 @@ def all_gather(self, tensor: torch.Tensor, group: Optional[Any] = None, sync_gra
         Return:
             A tensor of shape (world_size, batch, ...)
         """
-        return xm.all_gather(tensor.unsqueeze(0))
+        if isinstance(tensor, torch.Tensor) and tensor.dim() == 0:
+            tensor = tensor.unsqueeze(0)
+        return xm.all_gather(tensor)

pytorch_lightning/utilities/__init__.py

@@ -42,12 +42,10 @@
     _TORCH_QUANTIZE_AVAILABLE,
     _TORCHTEXT_AVAILABLE,
     _TORCHVISION_AVAILABLE,
+    _TPU_AVAILABLE,
     _XLA_AVAILABLE,
 )
 from pytorch_lightning.utilities.parsing import AttributeDict, flatten_dict, is_picklable  # noqa: F401
-from pytorch_lightning.utilities.xla_device import XLADeviceUtils  # noqa: F401
-
-_TPU_AVAILABLE = XLADeviceUtils.tpu_device_exists()
 
 FLOAT16_EPSILON = numpy.finfo(numpy.float16).eps
 FLOAT32_EPSILON = numpy.finfo(numpy.float32).eps

pytorch_lightning/utilities/distributed.py

@@ -20,7 +20,10 @@
 
 import torch
 
-log = logging.getLogger(__name__)
+from pytorch_lightning.utilities.imports import _TPU_AVAILABLE
+
+if _TPU_AVAILABLE:
+    import torch_xla.core.xla_model as xm
 
 if torch.distributed.is_available():
     from torch.distributed import group, ReduceOp
@@ -34,6 +37,9 @@ class group:
         WORLD = None
 
 
+log = logging.getLogger(__name__)
+
+
 def rank_zero_only(fn):
 
     @wraps(fn)
@@ -222,3 +228,9 @@ def all_gather_ddp_if_available(
             with torch.no_grad():
                 return AllGatherGrad.apply(tensor, group)
     return tensor
+
+
+def tpu_distributed() -> bool:
+    if _TPU_AVAILABLE:
+        return xm.xrt_world_size() > 1
+    return False

pytorch_lightning/utilities/imports.py

@@ -82,3 +82,7 @@ def _compare_version(package: str, op, version) -> bool:
 _TORCHTEXT_AVAILABLE = _module_available("torchtext")
 _TORCHVISION_AVAILABLE = _module_available('torchvision')
 _XLA_AVAILABLE = _module_available("torch_xla")
+
+from pytorch_lightning.utilities.xla_device import XLADeviceUtils  # noqa: E402
+
+_TPU_AVAILABLE = XLADeviceUtils.tpu_device_exists()

tests/models/test_tpu.py

@@ -16,13 +16,15 @@
 from unittest import mock
 
 import pytest
+import torch
 from torch.utils.data import DataLoader
 
 import tests.helpers.pipelines as tpipes
 import tests.helpers.utils as tutils
 from pytorch_lightning import Trainer
 from pytorch_lightning.accelerators import TPUAccelerator
 from pytorch_lightning.callbacks import EarlyStopping
+from pytorch_lightning.core.step_result import Result
 from pytorch_lightning.plugins import TPUSpawnPlugin
 from pytorch_lightning.trainer.states import TrainerState
 from pytorch_lightning.utilities import _TPU_AVAILABLE
@@ -397,3 +399,26 @@ def test_if_test_works_with_checkpoint_false(tmpdir):
     trainer = Trainer(max_epochs=1, tpu_cores=8, default_root_dir=tmpdir, fast_dev_run=True, checkpoint_callback=False)
     trainer.fit(model)
     assert trainer.state == TrainerState.FINISHED, f"Training failed with {trainer.state}"
+
+
+@RunIf(tpu=True)
+@pl_multi_process_test
+def test_tpu_sync_dist():
+    """Test tpu spawn sync dist operation """
+
+    def test_sync_dist(rank):
+        tensor = torch.tensor([1.0])
+        training_type_plugin = TPUSpawnPlugin()
+
+        res = Result()
+        res.log(
+            "test_tensor",
+            tensor,
+            sync_fn=training_type_plugin.reduce,
+            sync_dist=True,
+            sync_dist_op=torch.distributed.ReduceOp.SUM
+        )
+
+        assert res["test_tensor"].item() == 8, "Result-Log does not work properly with TPU Spawn and Tensors"
+
+    xmp.spawn(test_sync_dist, nprocs=8, start_method='fork')