Continuation of huggingface#942: additional float64 failure (huggingface#996)

* Add failing test for huggingface#940.

* Do not use torch.float64 in mps.

* style

* Temporarily skip add_noise for IPNDMScheduler.

Until huggingface#990 is addressed.

* Fix additional float64 error in mps.

* Improve add_noise test

* Slight edit – I think it's clearer this way.

Author: pcuenca

src/diffusers/schedulers/scheduling_lms_discrete.py

@@ -252,9 +252,13 @@ def add_noise(
     ) -> torch.FloatTensor:
         # Make sure sigmas and timesteps have the same device and dtype as original_samples
         self.sigmas = self.sigmas.to(device=original_samples.device, dtype=original_samples.dtype)
-        dtype = torch.float32 if original_samples.device.type == "mps" else timesteps.dtype
-        self.timesteps = self.timesteps.to(original_samples.device, dtype=dtype)
-        timesteps = timesteps.to(original_samples.device)
+        if original_samples.device.type == "mps" and torch.is_floating_point(timesteps):
+            # mps does not support float64
+            self.timesteps = self.timesteps.to(original_samples.device, dtype=torch.float32)
+            timesteps = timesteps.to(original_samples.device, dtype=torch.float32)
+        else:
+            self.timesteps = self.timesteps.to(original_samples.device)
+            timesteps = timesteps.to(original_samples.device)
 
         schedule_timesteps = self.timesteps
 

tests/test_scheduler.py

@@ -266,13 +266,14 @@ def test_add_noise_device(self):
                 continue
             scheduler_config = self.get_scheduler_config()
             scheduler = scheduler_class(**scheduler_config)
+            scheduler.set_timesteps(100)
 
             sample = self.dummy_sample.to(torch_device)
             scaled_sample = scheduler.scale_model_input(sample, 0.0)
             self.assertEqual(sample.shape, scaled_sample.shape)
 
             noise = torch.randn_like(scaled_sample).to(torch_device)
-            t = torch.tensor([10]).to(torch_device)
+            t = scheduler.timesteps[5][None]
             noised = scheduler.add_noise(scaled_sample, noise, t)
             self.assertEqual(noised.shape, scaled_sample.shape)