enable 2 test cases on XPU (huggingface#11332)

* enable 2 test cases on XPU

Signed-off-by: YAO Matrix <[email protected]>

* Apply style fixes

---------

Signed-off-by: YAO Matrix <[email protected]>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
Co-authored-by: Dhruv Nair <[email protected]>

Author: 3 people

tests/quantization/bnb/test_mixed_int8.py

@@ -523,13 +523,15 @@ def test_pipeline_cuda_placement_works_with_mixed_int8(self):
             torch_dtype=torch.float16,
             device_map=torch_device,
         )
+
         # CUDA device placement works.
+        device = torch_device if torch_device != "rocm" else "cuda"
         pipeline_8bit = DiffusionPipeline.from_pretrained(
             self.model_name,
             transformer=transformer_8bit,
             text_encoder_3=text_encoder_3_8bit,
             torch_dtype=torch.float16,
-        ).to("cuda")
+        ).to(device)
 
         # Check if inference works.
         _ = pipeline_8bit("table", max_sequence_length=20, num_inference_steps=2)

tests/quantization/utils.py

@@ -1,4 +1,10 @@
 from diffusers.utils import is_torch_available
+from diffusers.utils.testing_utils import (
+    backend_empty_cache,
+    backend_max_memory_allocated,
+    backend_reset_peak_memory_stats,
+    torch_device,
+)
 
 
 if is_torch_available():
@@ -30,9 +36,9 @@ def forward(self, input, *args, **kwargs):
     @torch.no_grad()
     @torch.inference_mode()
     def get_memory_consumption_stat(model, inputs):
-        torch.cuda.reset_peak_memory_stats()
-        torch.cuda.empty_cache()
+        backend_reset_peak_memory_stats(torch_device)
+        backend_empty_cache(torch_device)
 
         model(**inputs)
-        max_memory_mem_allocated = torch.cuda.max_memory_allocated()
-        return max_memory_mem_allocated
+        max_mem_allocated = backend_max_memory_allocated(torch_device)
+        return max_mem_allocated