update

app.py

@@ -1,7 +1,6 @@
 import os
 import shutil
 import tempfile
-import time
 import traceback
 import uuid
 from pathlib import Path
@@ -314,35 +313,22 @@ def run_demo(
     )
     try:
         # --- GPU-only inference (consumes ZeroGPU quota) ---
-        # Retry on transient "GPU task aborted" errors (common on ZeroGPU)
-        max_gpu_retries = 2
-        gpu_result = None
-        for attempt in range(max_gpu_retries + 1):
-            try:
-                gpu_result = run_demo_gpu(
-                    image=image,
-                    depth_file=depth_file,
-                    model_type=model_type,
-                    input_size=input_size,
-                    output_resolution_mode=output_resolution_mode,
-                    upsample_ratio=upsample_ratio,
-                    fx_org=fx_org,
-                    fy_org=fy_org,
-                    cx_org=cx_org,
-                    cy_org=cy_org,
-                    trace_path=trace_path,
-                )
-                break  # success
-            except Exception as gpu_exc:
-                is_aborted = "GPU task aborted" in str(gpu_exc)
-                if is_aborted and attempt < max_gpu_retries:
-                    Log.warning(
-                        f"[{request_id}] GPU task aborted (attempt {attempt + 1}/{max_gpu_retries + 1}), retrying..."
-                    )
-                    _append_trace(trace_path, f"ui:retry_{attempt + 1}_after_gpu_abort")
-                    time.sleep(2)
-                    continue
-                raise  # not retryable or out of retries
+        # ZeroGPU proxy tokens are bound to the current Gradio request.
+        # Retrying a @spaces.GPU call inside the same request can turn a transient
+        # "GPU task aborted" into a deterministic "Expired ZeroGPU proxy token".
+        gpu_result = run_demo_gpu(
+            image=image,
+            depth_file=depth_file,
+            model_type=model_type,
+            input_size=input_size,
+            output_resolution_mode=output_resolution_mode,
+            upsample_ratio=upsample_ratio,
+            fx_org=fx_org,
+            fy_org=fy_org,
+            cx_org=cx_org,
+            cy_org=cy_org,
+            trace_path=trace_path,
+        )
         _append_trace(trace_path, "ui:gpu_done, starting cpu postprocess")
 
         # --- CPU post-processing (no GPU quota consumed) ---
@@ -399,7 +385,7 @@ def run_demo(
                 "This is a HuggingFace ZeroGPU scheduling issue, not an inference bug.\n"
                 "Possible causes:\n"
                 "  - GPU quota exhausted (wait for quota to reset)\n"
-                "  - GPU task was preempted/aborted (try again)\n"
+                "  - GPU task was preempted/aborted (click the button again)\n"
                 "  - duration too high for remaining quota"
             )
         else:
@@ -499,9 +485,6 @@ with gr.Blocks(title="InfiniDepth Demo", theme=gr.themes.Soft(), css=CUSTOM_CSS,
 
 
     run_button.click(
-        fn=lambda: (None, None, [], "Running..."),
-        outputs=[depth_output, pcd_viewer, files_output, status],
-    ).then(
         fn=run_demo,
         inputs=[
             image_input,