Fix scheduler inference steps error with power of 3

src/diffusers/schedulers/scheduling_ddim.py

@@ -145,9 +145,10 @@ def set_timesteps(self, num_inference_steps: int, offset: int = 0):
             offset (`int`): TODO
         """
         self.num_inference_steps = num_inference_steps
-        self.timesteps = np.arange(
-            0, self.config.num_train_timesteps, self.config.num_train_timesteps // self.num_inference_steps
-        )[::-1].copy()
+        step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+        # creates integer timesteps by multiplying by ratio
+        # casting to int to avoid issues when num_inference_step is power of 3
+        self.timesteps = (np.arange(0, num_inference_steps) * step_ratio).astype(int)[::-1].copy()
         self.timesteps += offset
         self.set_format(tensor_format=self.tensor_format)
 

src/diffusers/schedulers/scheduling_pndm.py

@@ -141,9 +141,10 @@ def set_timesteps(self, num_inference_steps: int, offset: int = 0) -> torch.Floa
             offset (`int`): TODO
         """
         self.num_inference_steps = num_inference_steps
-        self._timesteps = list(
-            range(0, self.config.num_train_timesteps, self.config.num_train_timesteps // num_inference_steps)
-        )
+        step_ratio = self.config.num_train_timesteps // self.num_inference_steps
+        # creates integer timesteps by multiplying by ratio
+        # casting to int to avoid issues when num_inference_step is power of 3
+        self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).astype(int).tolist()
         self._offset = offset
         self._timesteps = np.array([t + self._offset for t in self._timesteps])
 

tests/test_scheduler.py

@@ -378,7 +378,7 @@ def test_time_indices(self):
 
     def test_inference_steps(self):
         for t, num_inference_steps in zip([1, 10, 50], [10, 50, 500]):
-            self.check_over_forward(num_inference_steps=num_inference_steps)
+            self.check_over_forward(time_step=t, num_inference_steps=num_inference_steps)
 
     def test_eta(self):
         for t, eta in zip([1, 10, 49], [0.0, 0.5, 1.0]):
@@ -621,6 +621,21 @@ def test_inference_steps(self):
         for t, num_inference_steps in zip([1, 5, 10], [10, 50, 100]):
             self.check_over_forward(time_step=t, num_inference_steps=num_inference_steps)
 
+    def test_pow_of_3_inference_steps(self):
+        num_inference_steps = 27
+
+        for scheduler_class in self.scheduler_classes:
+            sample = self.dummy_sample
+            residual = 0.1 * sample
+
+            scheduler_config = self.get_scheduler_config()
+            scheduler = scheduler_class(**scheduler_config)
+
+            scheduler.set_timesteps(num_inference_steps)
+
+            for i, t in enumerate(scheduler.prk_timesteps):
+                sample = scheduler.step_prk(residual, t, sample).prev_sample
+
     def test_inference_plms_no_past_residuals(self):
         with self.assertRaises(ValueError):
             scheduler_class = self.scheduler_classes[0]