fix: fallback to CPU when CUDA is not available

Previously, the code unconditionally attempted to move the model to the CUDA device (`self.model.to("cuda")`), which caused a runtime crash on systems where CUDA is not available (e.g., Apple M1/M2 or CPU-only environments). This resulted in the error:

AssertionError: Torch not compiled with CUDA enabled

The fix introduces a dynamic device selection:

    device = "cuda" if torch.cuda.is_available() else "cpu"
    self.model.to(device)

This change ensures compatibility across platforms and prevents crashes due to unavailable CUDA devices.

chat.py

@@ -93,7 +93,8 @@ class DOLPHIN:
             ckpt = try_rename_lagacy_weights(ckpt)
             self.model.load_state_dict(ckpt, strict=True)
 
-        self.model.to("cuda")
+        device = "cuda" if torch.cuda.is_available() else "cpu"
+        self.model.to(device)
         self.model.eval()
         transform_args = {
             "input_size": self.swin_args["img_size"],