Fix: Force Bfloat16 on CPU to satisfy PyTorch autocast requirements

ylff/utils/model_loader.py

@@ -236,10 +236,11 @@ def load_da3_model(
                 device = "cpu"
                 
                 # Monkeypatch torch.cuda.is_bf16_supported to avoid initializing CUDA on CPU machines
-                # This prevents "Found no NVIDIA driver" error in depth_anything_3/api.py
+                # AND return True to force usage of bfloat16, which is required for CPU autocast
+                # (PyTorch CPU AMP does not support float16, only bfloat16)
                 if hasattr(torch.cuda, "is_bf16_supported"):
-                    logger.info("Patching torch.cuda.is_bf16_supported to prevent CUDA initialization on CPU")
-                    torch.cuda.is_bf16_supported = lambda: False
+                    logger.info("Patching torch.cuda.is_bf16_supported=True to enable CPU bfloat16 autocast")
+                    torch.cuda.is_bf16_supported = lambda: True
 
         logger.info(f"Loading DA3 model: {model_name} on {device}")
         model = DepthAnything3.from_pretrained(model_name)