Do not use torch.long in mps (huggingface#1488)

* Do not use torch.long in mps

Addresses huggingface#1056.

* Use torch.int instead of float.

* Propagate changes.

* Do not silently change float -> int.

* Propagate changes.

* Apply suggestions from code review

Co-authored-by: Anton Lozhkov <[email protected]>

Co-authored-by: Anton Lozhkov <[email protected]>

Author: pcuenca

src/diffusers/models/unet_2d_condition.py

@@ -299,8 +299,14 @@ def forward(
         timesteps = timestep
         if not torch.is_tensor(timesteps):
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
-            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
-        elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if torch.is_floating_point(timesteps):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
 
         # broadcast to batch dimension in a way that's compatible with ONNX/Core ML

src/diffusers/pipelines/versatile_diffusion/modeling_text_unet.py

@@ -377,8 +377,14 @@ def forward(
         timesteps = timestep
         if not torch.is_tensor(timesteps):
             # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
-            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
-        elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if torch.is_floating_point(timesteps):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
             timesteps = timesteps[None].to(sample.device)
 
         # broadcast to batch dimension in a way that's compatible with ONNX/Core ML