[tests] use `torch_device` instead of `auto` for model testing (#29531)

* use torch_device

* skip for XPU

* Update tests/generation/test_utils.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

tests/generation/test_utils.py

@@ -1073,6 +1073,9 @@ class GenerationTesterMixin:
     @require_torch_multi_accelerator
     def test_model_parallel_beam_search(self):
         for model_class in self.all_generative_model_classes:
+            if "xpu" in torch_device:
+                return unittest.skip("device_map='auto' does not work with XPU devices")
+
             if model_class._no_split_modules is None:
                 continue