Scheduling fixes on MPS (#10549)

* use np.int32 in scheduling

* test_add_noise_device

* -np.int32, fixes

Author: hlky

src/diffusers/schedulers/scheduling_heun_discrete.py

@@ -342,7 +342,7 @@ def set_timesteps(
         timesteps = torch.from_numpy(timesteps)
         timesteps = torch.cat([timesteps[:1], timesteps[1:].repeat_interleave(2)])
 
-        self.timesteps = timesteps.to(device=device)
+        self.timesteps = timesteps.to(device=device, dtype=torch.float32)
 
         # empty dt and derivative
         self.prev_derivative = None

src/diffusers/schedulers/scheduling_lms_discrete.py

@@ -311,7 +311,7 @@ def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.devic
         sigmas = np.concatenate([sigmas, [0.0]]).astype(np.float32)
 
         self.sigmas = torch.from_numpy(sigmas).to(device=device)
-        self.timesteps = torch.from_numpy(timesteps).to(device=device)
+        self.timesteps = torch.from_numpy(timesteps).to(device=device, dtype=torch.float32)
         self._step_index = None
         self._begin_index = None
         self.sigmas = self.sigmas.to("cpu")  # to avoid too much CPU/GPU communication

tests/schedulers/test_scheduler_lcm.py

@@ -99,7 +99,7 @@ def test_add_noise_device(self, num_inference_steps=10):
             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)
+            noise = torch.randn(scaled_sample.shape).to(torch_device)
             t = scheduler.timesteps[5][None]
             noised = scheduler.add_noise(scaled_sample, noise, t)
             self.assertEqual(noised.shape, scaled_sample.shape)

tests/schedulers/test_schedulers.py

@@ -361,7 +361,7 @@ def model(sample, t, *args):
             if isinstance(t, torch.Tensor):
                 num_dims = len(sample.shape)
                 # pad t with 1s to match num_dims
-                t = t.reshape(-1, *(1,) * (num_dims - 1)).to(sample.device).to(sample.dtype)
+                t = t.reshape(-1, *(1,) * (num_dims - 1)).to(sample.device, dtype=sample.dtype)
 
             return sample * t / (t + 1)
 
@@ -722,7 +722,7 @@ def test_add_noise_device(self):
                 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)
+            noise = torch.randn(scaled_sample.shape).to(torch_device)
             t = scheduler.timesteps[5][None]
             noised = scheduler.add_noise(scaled_sample, noise, t)
             self.assertEqual(noised.shape, scaled_sample.shape)