A question about Qwen-3-vl-reranker for the difference between sglang inference and transformer inference.

[Serendipity-00]

I have viewed the GenerateReqInput() function in sglang/srt/entrypoints/openai/serving_rerank.py, the implement of the choice for YES and NO token is to choose the top-50 and then find the ID, other than just find the ID from the vocab, which may leads to score vanish to 0.0 or 1.0 if One of the token is not found.
But in other implements, (e.g. ms-swift for qwen3-vl-reranker) it is to choose the YES and NO token directly so that there is always an exact score.
I would like to know why this is the choice of sglang and how the number 50 is decided.
Thx a lot.

In sglang:

gen_request = GenerateReqInput(
    text=prompt,
    image_data=image_data if image_data else None,
    sampling_params={
        "max_new_tokens": 1,       
        "temperature": 0,         
    },
    return_logprob=True,
    top_logprobs_num=50,          
)

In ms-swift:

def get_generative_reranker_logits(lm_head_weight, tokenizer, hidden_states):
    positive_token = os.environ.get('GENERATIVE_RERANKER_POSITIVE_TOKEN', 'yes')
    negative_token = os.environ.get('GENERATIVE_RERANKER_NEGATIVE_TOKEN', 'no')
    positive_token_id = tokenizer.convert_tokens_to_ids(positive_token)
    negative_token_id = tokenizer.convert_tokens_to_ids(negative_token)
    weight = lm_head_weight[[positive_token_id, negative_token_id]]
    logits = F.linear(hidden_states, weight)
    return logits[..., 0:1] - logits[..., 1:2]