Fix log upload dropped by moving _spawn_log outside @spaces.GPU scope

ZeroGPU runs @spaces.GPU-decorated functions in an isolated subprocess;
when the function returns the subprocess exits, killing the daemon log
thread before the HuggingFace upload could complete. Moving _spawn_log
to infer() (main process) ensures the thread outlives the GPU subprocess.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app.py

@@ -339,7 +339,17 @@ def infer(images_b64_json, prompt, seed, randomize_seed, guidance_scale, steps,
     _validate_infer_inputs(pil_images, prompt)
     seed = _resolve_seed(seed, randomize_seed)
     width, height = update_dimensions_on_upload(pil_images[0], max_dim_for_mode(mode))
-    return _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height)
+    t0 = time.perf_counter()
+    try:
+        result_image, seed, duration = _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height)
+        # _spawn_log is called here (main process) so the thread survives after _infer_gpu's
+        # @spaces.GPU subprocess exits — previously the daemon thread was killed on subprocess exit.
+        _spawn_log(pil_images, result_image, prompt, seed, steps, guidance_scale, width, height, duration, True)
+        return result_image, seed
+    except Exception as e:
+        duration = time.perf_counter() - t0
+        _spawn_log(pil_images, None, prompt, seed, steps, guidance_scale, width, height, duration, False, str(e))
+        raise
 
 
 @spaces.GPU(duration=120)
@@ -392,8 +402,7 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height):
         print(f"[infer] VAE decode + postprocess done — {_gpu_mem_str(_cuda_ok, sync=True)} | t={time.perf_counter()-t0:.1f}s")
         timer.print_timings()
         duration = timer.elapsed_ms("pipe_start", "pipe_end") / 1000.0
-        _spawn_log(pil_images, result_image, prompt, seed, steps, guidance_scale, width, height, duration, True)
-        return result_image, seed
+        return result_image, seed, duration
     except Exception as e:
         print(f"[infer] ERROR: {type(e).__name__}: {e} | t={time.perf_counter()-t0:.1f}s")
         print(traceback.format_exc())
@@ -402,9 +411,7 @@ def _infer_gpu(pil_images, prompt, seed, guidance_scale, steps, width, height):
         except Exception as cuda_err:
             print(f"[infer] CUDA synchronize after error: {cuda_err}")
         timer.print_timings()
-        duration = time.perf_counter() - timer.wall_start("pipe_start")
-        _spawn_log(pil_images, None, prompt, seed, steps, guidance_scale, width, height, duration, False, str(e))
-        raise e
+        raise
     finally:
         gc.collect()
         torch.cuda.empty_cache()