Fix double precision casting complex buffers

CHANGELOG.md

@@ -342,6 +342,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed a bug where calling `log` with a `Metric` instance would raise an error if it was a nested attribute of the model ([#8181](https://github.com/PyTorchLightning/pytorch-lightning/pull/8181))
 
+
+- Fixed a bug where using `precision=64` would cause buffers with complex dtype to be cast to real ([#8208](https://github.com/PyTorchLightning/pytorch-lightning/pull/8208))
+
 ## [1.3.7] - 2021-06-22
 
 - Fixed a bug where skipping an optimizer while using amp causes amp to trigger an assertion error ([#7975](https://github.com/PyTorchLightning/pytorch-lightning/pull/7975))

pytorch_lightning/plugins/precision/double.py

@@ -96,7 +96,7 @@ def connect(
         incoming floating point data to double (``torch.float64``) precision. Does not alter `optimizers` or
         `lr_schedulers`.
         """
-        model = cast(pl.LightningModule, model.to(dtype=torch.float64))
+        model = cast(pl.LightningModule, model.double())
         model = LightningDoublePrecisionModule(model)
 
         return super().connect(model, optimizers, lr_schedulers)

tests/plugins/test_double_plugin.py

@@ -40,6 +40,11 @@ def __len__(self):
 
 class DoublePrecisionBoringModel(BoringModel):
 
+    def __init__(self):
+        super().__init__()
+
+        self.register_buffer("complex_buffer", torch.complex(torch.rand(10), torch.rand(10)), False)
+
     def training_step(self, batch, batch_idx):
         float_data, int_data = batch
         assert torch.tensor([0.]).dtype == torch.float64
@@ -77,9 +82,11 @@ def predict_step(self, batch, batch_idx, dataloader_idx=None):
 
     def on_fit_start(self):
         assert self.layer.weight.dtype == torch.float64
+        assert self.complex_buffer.dtype == torch.complex64
 
     def on_after_backward(self):
         assert self.layer.weight.grad.dtype == torch.float64
+        assert self.complex_buffer.dtype == torch.complex64
 
     def train_dataloader(self):
         dataset = RandomFloatIntDataset(32, 64)