You don’t need to provide the parameters to run inference. It should work like a charm. Padding is not required for inference.
But if you still want to set the max_length to 256 you can modify the tokenizer_config to 256, example: tokenizer_config.json · distilbert-base-uncased-finetuned-sst-2-english at main