feat: improve timings

app.py

@@ -491,11 +491,16 @@ def predict(
 
     # move to GPU only once the ZeroGPU context is active
     device = "cuda" if torch.cuda.is_available() else "cpu"
-    pipe.model.to(device)
+    # check where the model currently lives
+    current = next(pipe.model.parameters()).device.type  # "cpu" or "cuda"
+    # only move if needed
+    if current != device:
+        pipe.model.to(device)
+    tprint(f"model.ensure_on({device})")
+
     pipe.model.eval()
     print(f"Running on {next(pipe.model.parameters()).device}")
-
-    tprint(f"pipe.model.to({device})")
+    tprint("model ready to run inference")
 
     # run inference
     out = pipe(inputs)
@@ -597,6 +602,7 @@ def predict(
     fig.axes[-1].set_xlabel(region)
 
     png_path = _save_fig_png(fig)
+    tprint("figure png saved")
 
     meta = {
         "model_id": current_model_id,