perf: update dtype to bfloat16

config.json

@@ -2,7 +2,7 @@
   "architectures": [
     "Qwen3_5ForConditionalGeneration"
   ],
-  "dtype": "float32",
+  "dtype": "bfloat16",
   "image_token_id": 248056,
   "model_type": "qwen3_5",
   "text_config": {
@@ -10,7 +10,7 @@
     "attention_dropout": 0.0,
     "attn_output_gate": true,
     "bos_token_id": null,
-    "dtype": "float32",
+    "dtype": "bfloat16",
     "eos_token_id": 248044,
     "full_attention_interval": 4,
     "head_dim": 256,
@@ -83,7 +83,7 @@
   "vision_config": {
     "deepstack_visual_indexes": [],
     "depth": 24,
-    "dtype": "float32",
+    "dtype": "bfloat16",
     "hidden_act": "gelu_pytorch_tanh",
     "hidden_size": 1024,
     "in_channels": 3,
@@ -99,4 +99,4 @@
   },
   "vision_end_token_id": 248054,
   "vision_start_token_id": 248053
-}
+}