Fix Nightly AudioLDM2PipelineFastTests (#10556)

* Fix Nightly AudioLDM2PipelineFastTests

* add phonemizer to setup extras test

* fix

* make style

setup.py

@@ -135,6 +135,7 @@
     "transformers>=4.41.2",
     "urllib3<=2.0.0",
     "black",
+    "phonemizer",
 ]
 
 # this is a lookup table with items like:
@@ -227,6 +228,7 @@ def run(self):
     "scipy",
     "torchvision",
     "transformers",
+    "phonemizer",
 )
 extras["torch"] = deps_list("torch", "accelerate")
 

src/diffusers/dependency_versions_table.py

@@ -43,4 +43,5 @@
     "transformers": "transformers>=4.41.2",
     "urllib3": "urllib3<=2.0.0",
     "black": "black",
+    "phonemizer": "phonemizer",
 }

src/diffusers/pipelines/audioldm2/pipeline_audioldm2.py

@@ -237,7 +237,7 @@ def disable_vae_slicing(self):
         """
         self.vae.disable_slicing()
 
-    def enable_model_cpu_offload(self, gpu_id=0):
+    def enable_model_cpu_offload(self, gpu_id: Optional[int] = None, device: Union[torch.device, str] = "cuda"):
         r"""
         Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared
         to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward`
@@ -249,11 +249,23 @@ def enable_model_cpu_offload(self, gpu_id=0):
         else:
             raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.")
 
-        device = torch.device(f"cuda:{gpu_id}")
+        torch_device = torch.device(device)
+        device_index = torch_device.index
+
+        if gpu_id is not None and device_index is not None:
+            raise ValueError(
+                f"You have passed both `gpu_id`={gpu_id} and an index as part of the passed device `device`={device}"
+                f"Cannot pass both. Please make sure to either not define `gpu_id` or not pass the index as part of the device: `device`={torch_device.type}"
+            )
+
+        device_type = torch_device.type
+        device = torch.device(f"{device_type}:{gpu_id or torch_device.index}")
 
         if self.device.type != "cpu":
             self.to("cpu", silence_dtype_warnings=True)
-            torch.cuda.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
+            device_mod = getattr(torch, device.type, None)
+            if hasattr(device_mod, "empty_cache") and device_mod.is_available():
+                device_mod.empty_cache()  # otherwise we don't see the memory savings (but they probably exist)
 
         model_sequence = [
             self.text_encoder.text_model,

tests/pipelines/audioldm2/test_audioldm2.py

@@ -469,8 +469,8 @@ def test_xformers_attention_forwardGenerator_pass(self):
         pass
 
     def test_dict_tuple_outputs_equivalent(self):
-        # increase tolerance from 1e-4 -> 2e-4 to account for large composite model
-        super().test_dict_tuple_outputs_equivalent(expected_max_difference=2e-4)
+        # increase tolerance from 1e-4 -> 3e-4 to account for large composite model
+        super().test_dict_tuple_outputs_equivalent(expected_max_difference=3e-4)
 
     def test_inference_batch_single_identical(self):
         # increase tolerance from 1e-4 -> 2e-4 to account for large composite model