fixed TensorType for TPU device execution, limited the unit test chec… #3116
Conversation
…k fro torch::kDouble on infinity entries
test/cpp/test_aten_xla_tensor.cpp
Outdated
| AllClose(b, xla_b); | ||
|
|
||
| if (DebugUtil::ExperimentEnabled("nonzero")) { | ||
| if (DebugUtil::ExperimentEnabled("nonzero") && |
There was a problem hiding this comment.
Diff is still here, can you rebase again? bridge::AtenDeviceToXlaDevice(device).hw_type == DeviceType::TPU here has been removed from master. You might need to pull first.
There was a problem hiding this comment.
for some odd reason, the rebase and merge operations do not work as intended saying everything is up to date and showing the latest master commit present in this branch. Clearly something is wrong given this delta is still present. I ran a git checkout --patch on this file to resolve the problem.
torch_xla/csrc/aten_xla_type.cpp
Outdated
| auto element_type = TensorTypeToRawXlaType(self.scalar_type()); | ||
| XLATensor input_tensor = bridge::GetXlaTensor(self); | ||
| const Device& device = input_tensor.GetDevice(); | ||
| auto element_type = GetDevicePrimitiveType( |
There was a problem hiding this comment.
you can use MakeXlaPrimitiveType which takes scalar_type directly.
There was a problem hiding this comment.
Good shortcut. Done.
test/cpp/test_aten_xla_tensor.cpp
Outdated
| AllClose(b, xla_b); | ||
|
|
||
| if (DebugUtil::ExperimentEnabled("nonzero")) { | ||
| if (DebugUtil::ExperimentEnabled("nonzero") && |
There was a problem hiding this comment.
Diff is still here, can you rebase again? bridge::AtenDeviceToXlaDevice(device).hw_type == DeviceType::TPU here has been removed from master. You might need to pull first.
…a_tensor.cpp. improved variable use in torch_xla/csrc/aten_xla_type.cpp
|
Confirming this code passes on TPU tests. CC @JackCaoG |
Issue #3115