Fix converting only float type tensors in Lite (#10429)

* fix

* less code

* add test case

* add test cases

* update input

* add test cases

* add type hint

* add changelog note

Co-authored-by: Kaushik B <45285388+kaushikb11@users.noreply.github.com>

CHANGELOG.md

@@ -16,6 +16,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed an import error being caused by `PostLocalSGD` when `torch.distributed` not available ([#10359](https://github.com/PyTorchLightning/pytorch-lightning/pull/10359))
 - Fixed the logging with `on_step=True` in epoch-level hooks causing unintended side-effects. Logging with `on_step=True` in epoch-level hooks will now correctly raise an error ([#10409](https://github.com/PyTorchLightning/pytorch-lightning/pull/10409))
 - Fixed deadlocks for distributed training with `RichProgressBar` ([#10428](https://github.com/PyTorchLightning/pytorch-lightning/pull/10428))
+- Fixed an issue where the model wrapper in Lite converted non-floating point tensors to float ([#10429](https://github.com/PyTorchLightning/pytorch-lightning/pull/10429))
 
 
 ## [1.5.0] - 2021-11-02

pytorch_lightning/lite/wrappers.py

@@ -95,12 +95,17 @@ def forward(self, *args: Any, **kwargs: Any) -> Any:
         }
         # TODO (@awaelchli): let the precision plugin handle the conversion
         to_type = precision_to_type[precision]
-        args, kwargs = apply_to_collection([args, kwargs], function=lambda t: t.to(to_type), dtype=Tensor)
+
+        def _convert_float_tensor(t: Tensor) -> Tensor:
+            return t.to(to_type) if torch.is_floating_point(t) else t
+
+        args, kwargs = apply_to_collection([args, kwargs], function=_convert_float_tensor, dtype=Tensor)
 
         with self._precision_plugin.forward_context():
             output = self.module(*args, **kwargs)
 
-        output = apply_to_collection(output, function=lambda t: t.to(torch.get_default_dtype()), dtype=Tensor)
+        to_type = torch.get_default_dtype()
+        output = apply_to_collection(output, function=_convert_float_tensor, dtype=Tensor)
         return output
 
 

tests/lite/test_wrappers.py

@@ -40,8 +40,13 @@ def test_lite_module_wraps():
         (32, torch.float16, torch.float32),
         (32, torch.float32, torch.float32),
         (32, torch.float64, torch.float32),
+        (32, torch.int, torch.int),
         (16, torch.float32, torch.float16),
         (16, torch.float64, torch.float16),
+        (16, torch.long, torch.long),
+        pytest.param("bf16", torch.float32, torch.bfloat16, marks=RunIf(min_torch="1.10")),
+        pytest.param("bf16", torch.float64, torch.bfloat16, marks=RunIf(min_torch="1.10")),
+        pytest.param("bf16", torch.bool, torch.bool, marks=RunIf(min_torch="1.10")),
     ],
 )
 def test_lite_module_forward_conversion(precision, input_type, expected_type):
@@ -53,11 +58,11 @@ def check_autocast(forward_input):
         assert precision != 16 or torch.is_autocast_enabled()
         return forward_input
 
-    module = Mock(wraps=torch.nn.Linear(1, 1), side_effect=check_autocast)
+    module = Mock(wraps=torch.nn.Identity(), side_effect=check_autocast)
     lite_module = _LiteModule(module, lite._precision_plugin).to(device)
-    out = lite_module(torch.rand(1, dtype=input_type, device=device))
+    out = lite_module(torch.tensor([1, 2, 3], dtype=input_type, device=device))
     assert module.call_args[0][0].dtype == expected_type
-    assert out.dtype == torch.get_default_dtype()
+    assert out.dtype == input_type or out.dtype == torch.get_default_dtype()
 
 
 def test_lite_dataloader_iterator():