Revert model offloading (was not requested, slows computation)

gradio_app.py

@@ -142,11 +142,7 @@ def generate_masks_sam3(frames, text_prompt):
             else:
                 combined = np.zeros((H,W), dtype=np.uint8)
             raw_masks[model_out.frame_idx] = combined
-    result = [dilate_mask(raw_masks.get(i, np.zeros((H,W),dtype=np.uint8))) for i in range(len(frames))]
-    # Offload SAM3 to CPU after use to free VRAM
-    sam3_model.to("cpu")
-    torch.cuda.empty_cache()
-    return result
+    return [dilate_mask(raw_masks.get(i, np.zeros((H,W),dtype=np.uint8))) for i in range(len(frames))]
 
 # ── First-frame helpers ──
 
@@ -204,7 +200,6 @@ def infer(input_video, text_prompt, x1, y1, x2, y2, use_propainter):
         print(f"    BBox filter: x1={bx1} y1={by1} x2={bx2} y2={by2}")
 
     print(f"[2/6] SAM3 detecting '{text_prompt}'...")
-    sam3_model.to(device)  # move back to GPU for inference
     dilated_masks = generate_masks_sam3(frames, text_prompt.strip())
 
     if filter_bbox:
@@ -251,18 +246,10 @@ def infer(input_video, text_prompt, x1, y1, x2, y2, use_propainter):
     repaired_path = os.path.join(save_path, f"repaired_{ts}.mp4")
     t0 = time.time()
     if use_propainter:
-        propainter.fix_raft.to(device)
-        propainter.fix_flow_complete.to(device)
-        propainter.model.to(device)
         propainter.forward(crop_video_path, crop_mask_path, priori_path,
                            resize_ratio=1.0,
                            video_length=video_length, ref_stride=10,
                            neighbor_length=10, subvideo_length=50, mask_dilation=8)
-        # Offload ProPainter to CPU after use
-        propainter.fix_raft.to("cpu")
-        propainter.fix_flow_complete.to("cpu")
-        propainter.model.to("cpu")
-        torch.cuda.empty_cache()
     else:
         import shutil
         shutil.copy2(crop_video_path, priori_path)