Slow Response time in reasoning model parsing

[enochdongyi]

I try to apply xgrammar in JSON format structured decoding with Vllm, the outputs should have same structure like
<think> xxxxxxxxxx </think> ```json\n [{'Question': 'xxxxxxx', 'Answer': 'xxxxxxxxx'}]

Under my experiment with Qwen-deepseek-distill-32B, the total running time increased by 11 times(26min-4h30min)

from vllm.sampling_params import GuidedDecodingParams
tokenizer = llm.get_tokenizer() 
simplified_grammar =  r"""
root   ::= basic_string_1 ws think-pre  arr "\n```"
think-pre  ::= "</think>\n\n```json\n"
arr  ::=
  "[" (object (",\n" ws object){2}) "]"

object ::= "{\"question\": " ws basic_string ws ",\n\"answer\" :" ws basic_string ws "}" ws
basic_string ::= (([\"] basic_string_1 [\"]))
basic_string_1 ::= "" | [^"\\\x00-\x1F] basic_string_1 | "\\" escape basic_string_1
escape ::= ["\\/bfnrt] | "u" [A-Fa-f0-9] [A-Fa-f0-9] [A-Fa-f0-9] [A-Fa-f0-9]
# Optional space: by convention, applied in this grammar after literal chars when allowed
ws ::= [ \n\t]?
"""
guided_decoding_params = GuidedDecodingParams(grammar=simplified_grammar)
sampling_params2 = SamplingParams(n=3, temperature=0.8, top_p=1, repetition_penalty=1.05, max_tokens=2048,
                                  guided_decoding=guided_decoding_params)

outputs2 = llm.generate(
    input_prompts,
    sampling_params=sampling_params2,
)

Any thoughts on applying xgrammar in reasoning models? The low speed is caused by bad GBNF grammar writing, or structured decoding is hard for reasoning model