deepcopy model and processor

app.py

@@ -509,11 +509,10 @@ def on_image_click(
 @spaces.GPU()
 def propagate_masks(GLOBAL_STATE: gr.State):
     if GLOBAL_STATE is None or GLOBAL_STATE.inference_session is None:
-        # yield "Load a video first.", gr.update()
-        return GLOBAL_STATE, "Load a video first.", gr.update()
+        yield GLOBAL_STATE, "Load a video first.", gr.update()
 
-    processor = GLOBAL_STATE.processor
-    model = GLOBAL_STATE.model
+    processor = deepcopy(GLOBAL_STATE.processor)
+    model = deepcopy(GLOBAL_STATE.model)
     inference_session = deepcopy(GLOBAL_STATE.inference_session)
     # set inference device to cuda to use zero gpu
     inference_session.inference_device = "cuda"
@@ -524,7 +523,7 @@ def propagate_masks(GLOBAL_STATE: gr.State):
     processed = 0
 
     # Initial status; no slider change yet
-    # yield f"Propagating masks: {processed}/{total}", gr.update()
+    yield GLOBAL_STATE, f"Propagating masks: {processed}/{total}", gr.update()
 
     last_frame_idx = 0
     with torch.inference_mode():
@@ -547,25 +546,15 @@ def propagate_masks(GLOBAL_STATE: gr.State):
 
             processed += 1
             # Every 15th frame (or last), move slider to current frame to update preview via slider binding
-            # if processed % 15 == 0 or processed == total:
-            #     yield f"Propagating masks: {processed}/{total}", gr.update(value=frame_idx)
-            # else:
-            #     yield f"Propagating masks: {processed}/{total}", gr.update()
-
-    model.to("cpu")
-    inference_session.inference_device = "cpu"
-    inference_session.cache.inference_device = "cpu"
-    gc.collect()
-    torch.cuda.empty_cache()
+            if processed % 15 == 0 or processed == total:
+                yield GLOBAL_STATE, f"Propagating masks: {processed}/{total}", gr.update(value=frame_idx)
+            else:
+                yield GLOBAL_STATE, f"Propagating masks: {processed}/{total}", gr.update()
+
     text = f"Propagated masks across {processed} frames for {len(inference_session.obj_ids)} objects."
 
     # Final status; ensure slider points to last processed frame
-    # yield (
-    #     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, text, gr.update(value=last_frame_idx)
+    yield GLOBAL_STATE, text, gr.update(value=last_frame_idx)
 
 
 def reset_session(GLOBAL_STATE: gr.State) -> tuple[AppState, Image.Image, int, int, str]: