Device generic test framework + dtypes

Writing-tests-that-run-on-all-available-device-types.md

@@ -1,5 +1,7 @@
 PyTorch's test framework lets you write tests that can run on all available device types. These "device generic" tests are a great way to ensure operations run properly no matter what hardware PyTorch is using.
 
+### Writing a Device Generic Test
+
 To write a device generic test you have to do three things:
 
 1. Your test should accept two arguments, self and device. The latter will be a string designating a device, like 'cpu,' 'cuda,' or 'xla.' 
@@ -8,10 +10,52 @@ To write a device generic test you have to do three things:
 
 When the test suite is run it will instantiate a device-specific version of the device generic test class. `TestTorchDeviceType` becomes `TestTorchDeviceTypeCPU` and `TestTorchDeviceTypeCUDA`, for example. Its tests are added to each of these classes and have the device type appended to their name. `test_diagonal`, for example, becomes `test_diagonal_cpu` and `test_diagonal_cuda`, respectively. These tests are called with the appropriate device type string when run.
 
+In Python, a test suite like this
+
+```python
+TestTorchDeviceType(TestCase):
+def testX(self, device):
+...
+```
+
+is translated to
+
+```python
+TestTorchDeviceTypeCPU(TestCase):
+def testX_cpu(self, device='cpu'):
+...
+TestTorchDeviceTypeCUDA(TestCase):
+def testX_cuda(self, device='cuda'):
+```
+
 These tests can be run directly, `python test_torch.py TestTorchDeviceTypeCPU.test_diagonal_cpu`, or filtered using pytest. The command `pytest test_torch.py -k 'test_diagonal'` will run both `test_diagonal_cpu` and `test_diagonal_cuda`. 
 
 See `TestTorchDeviceType` in test_torch.py for examples.
 
+### Writing a Device Generic Test with Dtype Variants
+
+Tests can accept a third argument, 'dtype,' if they use the @dtypes decorator, like so:
+
+```python
+@dtypes(torch.half, torch.float, torch.double)
+testX(self, device, dtype)
+```
+
+This will instantiate variants of testX for each available device type and each specified dtype...
+
+```python
+testX_cpu_half(self, 'cpu', torch.half)
+testX_cpu_float(self, 'cpu', torch.float)
+testX_cpu_double(self, 'cpu', torch.double)
+testX_cuda_half(self, 'cuda', torch.half)
+...
+```
+
+These tests can be run just like device generic tests without dtypes.
+
+
+### Older Methods (Do Not Use)
+
 Please do not use older methods such as:
 
 1) `use_cuda` variants: