Set kv_cache_dtype + torch_dtype to float32 (matches actual ONNX weights)

config.json

@@ -16,7 +16,7 @@
   "conv_bias": false,
   "conv_dim": 2048,
   "conv_use_xavier_init": true,
-  "torch_dtype": "float16",
+  "torch_dtype": "float32",
   "eos_token_id": 7,
   "full_attn_idxs": null,
   "hidden_size": 2048,
@@ -63,7 +63,7 @@
       "model_q4.onnx": 1
     },
     "kv_cache_dtype": {
-      "q4": "float16"
+      "q4": "float32"
     }
   }
 }