use deepcopy on inference session instead

app.py

@@ -1,5 +1,6 @@
 import colorsys
 import gc
+from copy import deepcopy
 from typing import Optional
 
 import gradio as gr
@@ -513,7 +514,7 @@ def propagate_masks(GLOBAL_STATE: gr.State):
 
     processor = GLOBAL_STATE.processor
     model = GLOBAL_STATE.model
-    inference_session = GLOBAL_STATE.inference_session
+    inference_session = deepcopy(GLOBAL_STATE.inference_session)
     # set inference device to cuda to use zero gpu
     inference_session.inference_device = "cuda"
     inference_session.cache.inference_device = "cuda"
@@ -556,8 +557,7 @@ def propagate_masks(GLOBAL_STATE: gr.State):
     inference_session.cache.inference_device = "cpu"
     gc.collect()
     torch.cuda.empty_cache()
-    test = f"Propagated masks across {processed} frames for {len(inference_session.obj_ids)} objects."
-    inference_session.reset_inference_session()
+    text = f"Propagated masks across {processed} frames for {len(inference_session.obj_ids)} objects."
 
     # Final status; ensure slider points to last processed frame
     # yield (
@@ -565,11 +565,7 @@ def propagate_masks(GLOBAL_STATE: gr.State):
     #     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.",
-        gr.update(value=last_frame_idx),
-    )
+    return GLOBAL_STATE, text, gr.update(value=last_frame_idx)
 
 
 def reset_session(GLOBAL_STATE: gr.State) -> tuple[AppState, Image.Image, int, int, str]: