Llama models apparently adding extra input tokens.

[snova-rodrigom]

Hi, I've tested multiple Llama models 3.1-3.3 from SambaNova, Fireworks and TogetherAI and have noticed that additional tokens are apparently being added per API model call. I'm using the right model tokenizer to cut and count the number of desire tokens to process as prompt tokens, however APIs always return ~30 extra tokens more, which is not expected. Since this is happening in all of them, is this a model related issue? is there a way to get the exact amount of prompt tokens expected?