set device to cpu and remove empty cuda cache

app.py

@@ -130,7 +130,7 @@ class AppState:
         self.inference_session = None
         self.model: Optional[AutoModel] = None
         self.processor: Optional[Sam2VideoProcessor] = None
-        self.device: str = "cuda"
+        self.device: str = "cpu"
         self.dtype: torch.dtype = torch.bfloat16
         self.video_fps: float | None = None
         self.masks_by_frame: dict[int, dict[int, np.ndarray]] = {}
@@ -154,6 +154,9 @@ class AppState:
         self.model_repo_id: str | None = None
         self.session_repo_id: str | None = None
 
+    def __repr__(self):
+        return f"AppState(video_frames={self.video_frames}, inference_session={self.inference_session is not None}, model={self.model is not None}, processor={self.processor is not None}, device={self.device}, dtype={self.dtype}, video_fps={self.video_fps}, masks_by_frame={self.masks_by_frame}, color_by_obj={self.color_by_obj}, clicks_by_frame_obj={self.clicks_by_frame_obj}, boxes_by_frame_obj={self.boxes_by_frame_obj}, composited_frames={self.composited_frames}, current_frame_idx={self.current_frame_idx}, current_obj_id={self.current_obj_id}, current_label={self.current_label}, current_clear_old={self.current_clear_old}, current_prompt_type={self.current_prompt_type}, pending_box_start={self.pending_box_start}, pending_box_start_frame_idx={self.pending_box_start_frame_idx}, pending_box_start_obj_id={self.pending_box_start_obj_id}, is_switching_model={self.is_switching_model}, model_repo_key={self.model_repo_key}, model_repo_id={self.model_repo_id}, session_repo_id={self.session_repo_id})"
+
     @property
     def num_frames(self) -> int:
         return len(self.video_frames)
@@ -189,8 +192,6 @@ def load_model_if_needed(GLOBAL_STATE: gr.State) -> tuple[AutoModel, Sam2VideoPr
     print(f"Loading model from {desired_repo}")
     device, dtype = get_device_and_dtype()
     # free up the gpu memory
-    torch.cuda.empty_cache()
-    gc.collect()
     model = AutoModel.from_pretrained(desired_repo)
     processor = Sam2VideoProcessor.from_pretrained(desired_repo)
     model.to(device, dtype=dtype)
@@ -225,11 +226,6 @@ def ensure_session_for_current_model(GLOBAL_STATE: gr.State) -> None:
                 pass
             GLOBAL_STATE.inference_session = None
             gc.collect()
-            try:
-                if torch.cuda.is_available():
-                    torch.cuda.empty_cache()
-            except Exception:
-                pass
             GLOBAL_STATE.inference_session = processor.init_video_session(
                 video=GLOBAL_STATE.video_frames,
                 inference_device=device,
@@ -566,6 +562,7 @@ def propagate_masks(GLOBAL_STATE: gr.State):
     #     f"Propagated masks across {processed} frames for {len(inference_session.obj_ids)} objects.",
     #     gr.update(value=last_frame_idx),
     # )
+    print("global state", GLOBAL_STATE)
     return (
         GLOBAL_STATE,
         f"Propagated masks across {processed} frames for {len(inference_session.obj_ids)} objects.",
@@ -596,11 +593,6 @@ def reset_session(GLOBAL_STATE: gr.State) -> tuple[AppState, Image.Image, int, i
         pass
     GLOBAL_STATE.inference_session = None
     gc.collect()
-    try:
-        if torch.cuda.is_available():
-            torch.cuda.empty_cache()
-    except Exception:
-        pass
     ensure_session_for_current_model(GLOBAL_STATE)
 
     # Keep current slider index if possible