Fix the value of model_max_length

Make model_max_length value match that of the original model meta-llama/Llama-3.2-90B-Vision-Instruct

tokenizer_config.json

@@ -2065,7 +2065,7 @@
     "input_ids",
     "attention_mask"
   ],
-  "model_max_length": 512,
+  "model_max_length": 131072,
   "pad_token": "<|eot_id|>",
   "padding_side": "left",
   "tokenizer_class": "PreTrainedTokenizerFast"