Fix broken tests with new accelerate version (#873)

* Fix broken tests with new accelerate version

Latest changes in accelerate led to breaking tests. Unfortunately, the
solution is to use a privat method on AcceleratorState but the devs said
that this would be the best solution.

This PR also sets the requirements for accelerate to be v0.11 or higher.
This not only makes our code simpler, but also opens the door to
supporting gradient accumulation with accelerate (which will come in a
separate PR).

* Skip AMP tests if no CUDA device is found

requirements-dev.txt

@@ -1,4 +1,4 @@
-accelerate
+accelerate>=0.11.0
 fire
 flaky
 future>=0.17.1

skorch/tests/test_helper.py

@@ -798,7 +798,13 @@ def accelerator_cls(self):
         pytest.importorskip('accelerate')
 
         from accelerate import Accelerator
+        from accelerate.state import AcceleratorState
 
+        # We have to use this private method because otherwise, the
+        # AcceleratorState is not completely reset, which results in an error
+        # initializing the Accelerator more than once in the same process.
+        # pylint: disable=protected-access
+        AcceleratorState._reset_state()
         return Accelerator
 
     @pytest.fixture
@@ -821,16 +827,14 @@ def test_mixed_precision(self, net_cls, accelerator_cls, data, mixed_precision):
         # Only test if training works at all, no specific test of whether the
         # indicated precision is actually used, since that depends on the
         # underlying hardware.
-        import accelerate
+        from accelerate.utils import is_bf16_available
 
-        if LooseVersion(accelerate.__version__) > '0.5.1':
-            accelerator = accelerator_cls(mixed_precision=mixed_precision)
-        elif mixed_precision == 'bf16':
-            pytest.skip('bf16 only supported in accelerate version > 0.5.1')
-        else:
-            fp16 = mixed_precision == 'fp16'
-            accelerator = accelerator_cls(fp16=fp16)
+        if (mixed_precision != 'no') and not torch.cuda.is_available():
+            pytest.skip('skipping AMP test because device does not support it')
+        if (mixed_precision == 'bf16') and not is_bf16_available():
+            pytest.skip('skipping bf16 test because device does not support it')
 
+        accelerator = accelerator_cls(mixed_precision=mixed_precision)
         net = net_cls(accelerator=accelerator)
         X, y = data
         net.fit(X, y)  # does not raise