* Fix nondeterministic tests for GPU runs * force SD fast tests to the CPU
* Add torch_device to the VE pipeline * Mark the training test with slow