Change dtype to bf16

This enables vLLM to load the model in bf16 by default.

config.json

@@ -29,7 +29,7 @@
   },
   "rope_theta": 500000.0,
   "tie_word_embeddings": false,
-  "torch_dtype": "float32",
+  "torch_dtype": "bfloat16",
   "transformers_version": "4.46.3",
   "use_cache": true,
   "vocab_size": 128256