[SAM3] Fix MPS race condition in add_point_inputs (#43042)

[SAM2] Fix MPS race condition in add_point_inputs

Author: vykhovanets

src/transformers/models/edgetam_video/modeling_edgetam_video.py

@@ -1001,7 +1001,7 @@ def add_point_inputs(self, obj_idx: int, frame_idx: int, inputs: dict):
         device_inputs = {}
         for key, value in inputs.items():
             if isinstance(value, torch.Tensor):
-                device_inputs[key] = value.to(self.inference_device, non_blocking=True)
+                device_inputs[key] = value.to(self.inference_device, non_blocking=False)
             else:
                 device_inputs[key] = value
         self.point_inputs_per_obj[obj_idx][frame_idx] = device_inputs

src/transformers/models/sam2_video/modeling_sam2_video.py

@@ -209,7 +209,7 @@ def add_point_inputs(self, obj_idx: int, frame_idx: int, inputs: dict):
         device_inputs = {}
         for key, value in inputs.items():
             if isinstance(value, torch.Tensor):
-                device_inputs[key] = value.to(self.inference_device, non_blocking=True)
+                device_inputs[key] = value.to(self.inference_device, non_blocking=False)
             else:
                 device_inputs[key] = value
         self.point_inputs_per_obj[obj_idx][frame_idx] = device_inputs

src/transformers/models/sam2_video/modular_sam2_video.py

@@ -478,7 +478,7 @@ def add_point_inputs(self, obj_idx: int, frame_idx: int, inputs: dict):
         device_inputs = {}
         for key, value in inputs.items():
             if isinstance(value, torch.Tensor):
-                device_inputs[key] = value.to(self.inference_device, non_blocking=True)
+                device_inputs[key] = value.to(self.inference_device, non_blocking=False)
             else:
                 device_inputs[key] = value
         self.point_inputs_per_obj[obj_idx][frame_idx] = device_inputs

src/transformers/models/sam3_tracker_video/modeling_sam3_tracker_video.py

@@ -213,7 +213,7 @@ def add_point_inputs(self, obj_idx: int, frame_idx: int, inputs: dict):
         device_inputs = {}
         for key, value in inputs.items():
             if isinstance(value, torch.Tensor):
-                device_inputs[key] = value.to(self.inference_device, non_blocking=True)
+                device_inputs[key] = value.to(self.inference_device, non_blocking=False)
             else:
                 device_inputs[key] = value
         self.point_inputs_per_obj[obj_idx][frame_idx] = device_inputs