SDNQ uint4 JoyCaption — complete single-file safetensors (in-place CPU quant, 5.85 GB)

config.json

@@ -71,8 +71,8 @@
     "quant_method": "sdnq",
     "weights_dtype": "uint4",
     "group_size": 128,
-    "quantization_device": "cpu",
-    "return_device": "cpu",
+    "quantization_device": "cuda",
+    "return_device": "cuda",
     "use_svd": false,
     "quant_conv": false,
     "quant_embedding": false,