Focal native twin: fresh instance build instead of CUDA deepcopy

app.py

@@ -1,6 +1,5 @@
 """FocalCodec 12.5 Hz weight-quantization lab (Gradio / ZeroGPU)."""
 
-import copy
 import hashlib
 import time
 
@@ -36,7 +35,11 @@ BASE_STATE = {k: v.detach().cpu().clone() for k, v in codec.state_dict().items()
 SIZES = {level: core.model_size_mb(BASE_STATE, level) for level in QUANT_LEVELS}
 NATIVE_LEVEL = "int8-native"
 try:
-    codec_i8 = int8nn.apply_int8(copy.deepcopy(codec), BASE_STATE)  # W8A8 twin: int8-resident weights
+    # fresh instance, not deepcopy: a second copy of the live (CUDA) codec is not deep-copyable there
+    codec_i8 = int8nn.apply_int8(
+        torch.hub.load(HUB_REPO, "focalcodec", config=MODEL_REPO, trust_repo=True).eval().requires_grad_(False).cuda(),
+        BASE_STATE,
+    )
 except Exception as exc:  # noqa: BLE001  a runtime quirk must not take the demo down
     codec_i8 = None
     print(f"[int8] native FocalCodec disabled: {exc!r}")