I have a reranker (cross encoder) which relies on an absolute positional embedding, limited to 8192 tokens.
In a situation where the length of prompt + document > 8192 tokens, it stops.
Tried using kwargs but didnt find the one passed to the tokenizer

tokenizer = AutoTokenizer.from_pretrained(reranker_path)
model = AutoModelForSequenceClassification.from_pretrained(reranker_path, torch_dtype=torch.float16)
similarity_scorer = Similarity(
	(model, tokenizer), 
	crossencode=True, 
)

How to truncate on the right?