Do not pass non_blocking=True if it does not support this argument (L…

…ightning-AI#2910)

* add docs

* non blocking only on tensor

* changelog

* add test case

* add test comment

* update changelog


changelog


chlog

CHANGELOG.md

@@ -123,6 +123,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed save apex scaler states ([#2828](https://github.com/PyTorchLightning/pytorch-lightning/pull/2828))
 
+- Fixed passing `non_blocking=True` when transferring a batch object that does not support it ([#2910](https://github.com/PyTorchLightning/pytorch-lightning/pull/2910))
+
 ## [0.8.5] - 2020-07-09
 
 ### Added

pytorch_lightning/core/hooks.py

@@ -309,8 +309,13 @@ def transfer_batch_to_device(self, batch, device)
         Note:
             This hook should only transfer the data and not modify it, nor should it move the data to
             any other device than the one passed in as argument (unless you know what you are doing).
-            The :class:`~pytorch_lightning.trainer.trainer.Trainer` already takes care of splitting the
-            batch and determines the target devices.
+
+        Note:
+            This hook only runs on single GPU training (no data-parallel). If you need multi-GPU support
+            for your custom batch objects, you need to define your custom
+            :class:`~torch.nn.parallel.DistributedDataParallel` or
+            :class:`~pytorch_lightning.overrides.data_parallel.LightningDistributedDataParallel` and
+            override :meth:`~pytorch_lightning.core.lightning.LightningModule.configure_ddp`.
 
         See Also:
             - :func:`~pytorch_lightning.utilities.apply_func.move_data_to_device`

pytorch_lightning/utilities/apply_func.py

@@ -104,6 +104,7 @@ def batch_to(data):
                 setattr(device_data, field, device_field)
             return device_data
 
-        return data.to(device, non_blocking=True)
+        kwargs = dict(non_blocking=True) if isinstance(data, torch.Tensor) else {}
+        return data.to(device, **kwargs)
 
     return apply_to_collection(batch, dtype=(TransferableDataType, Batch), function=batch_to)

tests/models/test_gpu.py

@@ -1,4 +1,5 @@
 from collections import namedtuple
+from unittest.mock import patch
 
 import pytest
 import torch
@@ -384,3 +385,24 @@ def to(self, *args, **kwargs):
 
     assert batch.text.type() == 'torch.cuda.LongTensor'
     assert batch.label.type() == 'torch.cuda.LongTensor'
+
+
+@pytest.mark.skipif(not torch.cuda.is_available(), reason="test requires GPU machine")
+def test_non_blocking():
+    """ Tests that non_blocking=True only gets passed on torch.Tensor.to, but not on other objects. """
+    trainer = Trainer()
+
+    batch = torch.zeros(2, 3)
+    with patch.object(batch, 'to', wraps=batch.to) as mocked:
+        trainer.transfer_batch_to_gpu(batch, 0)
+        mocked.assert_called_with(torch.device('cuda', 0), non_blocking=True)
+
+    class BatchObject(object):
+
+        def to(self, *args, **kwargs):
+            pass
+
+    batch = BatchObject()
+    with patch.object(batch, 'to', wraps=batch.to) as mocked:
+        trainer.transfer_batch_to_gpu(batch, 0)
+        mocked.assert_called_with(torch.device('cuda', 0))