Remove CUDA post-processing outside the inference funciton

demo_gradio.py

@@ -88,11 +88,30 @@ def process_image_once(inputs, enable_mask):
         model.module.return_masks = enable_mask
         outputs, _, _, _, masks = model(img, bboxes)
 
-    outputs[0]["pred_boxes"] = outputs[0]["pred_boxes"].cpu()
-    outputs[0]["box_v"] = outputs[0]["box_v"].cpu()
-    masks[0] = masks[0].cpu() if enable_mask else None
+    # ------------------------------------------------------------------
+    # ZeroGPU requirement: return ONLY CPU-native objects to main process.
+    # Do NOT return CUDA tensors, and avoid returning output dicts that may
+    # contain additional CUDA tensors beyond pred_boxes/box_v.
+    # ------------------------------------------------------------------
+    out0 = outputs[0]
 
-    return image, outputs, masks, img, scale, drawn_boxes
+    pred_boxes_cpu = out0["pred_boxes"].detach().float().cpu()
+    box_v_cpu = out0["box_v"].detach().float().cpu()
+
+    outputs_cpu = [{
+        "pred_boxes": pred_boxes_cpu,
+        "box_v": box_v_cpu,
+    }]
+
+    if enable_mask and masks is not None and masks[0] is not None:
+        masks_cpu = [masks[0].detach().float().cpu()]
+    else:
+        masks_cpu = [None]
+
+    # img is only used for shape in post_process, so return a CPU tensor
+    img_cpu = img.detach().cpu()
+
+    return image, outputs_cpu, masks_cpu, img_cpu, float(scale), drawn_boxes
 
 
 # -----------------------------
@@ -188,7 +207,6 @@ def post_process(image, outputs, masks, img, scale, drawn_boxes, enable_mask, th
         draw.rectangle([x1, y1, x2, y2], outline=box_rgb, width=box_width)
 
     # --- Exemplar boxes (user-drawn): keep clear but unobtrusive, no text ---
-    # You can adjust these two colors if you want them to be more/less prominent.
     exemplar_outline = (255, 255, 255, 255)  # white
     exemplar_inner = (0, 0, 0, 255)          # black
     for box in drawn_boxes: