Protect numpy from bfloat16 errors on T4 and restrict model upcast to CPU only

app.py

@@ -64,7 +64,8 @@ if model_installed:
             
     model.load_state_dict(image_state_dict, strict=False)
     model.to(device)
-    model.to(torch.float32)  # Force upcast from checkpoint's native bfloat16 to float32 for CPU inference
+    if not torch.cuda.is_available():
+        model.to(torch.float32)  # Force upcast from checkpoint's native bfloat16 to float32 for CPU inference
     processor = Sam3Processor(model)
     if not torch.cuda.is_available():
         processor.device = "cpu"
@@ -120,8 +121,8 @@ def extract_assets(input_image):
             if masks is None or len(masks) == 0:
                 continue
                 
-            if torch.is_tensor(masks): masks = masks.cpu().numpy()
-            if torch.is_tensor(scores): scores = scores.cpu().numpy()
+            if torch.is_tensor(masks): masks = masks.float().cpu().numpy()
+            if torch.is_tensor(scores): scores = scores.float().cpu().numpy()
             
             for j in range(len(masks)):
                 m = masks[j]