# Structured Outputs

**URL:** https://discuss.vllm.ai/c/features/structured-outputs/17.md

[Latest](https://discuss.vllm.ai/latest.md) · [Categories](https://discuss.vllm.ai/categories.md)

---

## [About the Structured Outputs category](https://discuss.vllm.ai/t/about-the-structured-outputs-category/85)

<div class="topic-metadata">

**Author:** [@DarkLight1337](https://discuss.vllm.ai/u/DarkLight1337)\
**Replies:** 0

</div>

---

## [Does LMFE\_STRICT\_JSON\_FIELD\_ORDER not work?](https://discuss.vllm.ai/t/does-lmfe-strict-json-field-order-not-work/1874)

<div class="topic-metadata">

**Author:** [@endolith](https://discuss.vllm.ai/u/endolith)\
**Replies:** 2\
**Last updated:** [November 19, 2025, 9:59pm UTC](https://discuss.vllm.ai/t/does-lmfe-strict-json-field-order-not-work/1874 "2025-11-19T21:59:18Z")

</div>

I tried to enable LMFE\_STRICT\_JSON\_FIELD\_ORDER = true in the env vars in Runpod but it still allows the LLM to output json in whatever order it wants (such as outputting the actual response first and the reasoning behind…

---

## [General questions on structured output backend](https://discuss.vllm.ai/t/general-questions-on-structured-output-backend/1444)

<div class="topic-metadata">

**Author:** [@hustxiayang](https://discuss.vllm.ai/u/hustxiayang)\
**Replies:** 9\
**Last updated:** [September 3, 2025, 12:59am UTC](https://discuss.vllm.ai/t/general-questions-on-structured-output-backend/1444 "2025-09-03T00:59:51Z")

</div>

@RunLLM Could you please elaborate the efficiency comparison between different structured output backend(outlines, xgrammar, guidance), and also the features support difference? Thanks!

---

## [Specifying special tokens](https://discuss.vllm.ai/t/specifying-special-tokens/595)

<div class="topic-metadata">

**Author:** [@g-eoj](https://discuss.vllm.ai/u/g-eoj)\
**Replies:** 5\
**Last updated:** [May 8, 2025, 11:00pm UTC](https://discuss.vllm.ai/t/specifying-special-tokens/595 "2025-05-08T23:00:20Z")

</div>

Are there any known methods for “guiding” a model to output a special token (like \</think\>) rather than having it select tokens for substrings of the special token?

---

## [Grammar CPU bound performance](https://discuss.vllm.ai/t/grammar-cpu-bound-performance/372)

<div class="topic-metadata">

**Author:** [@supersteves](https://discuss.vllm.ai/u/supersteves)\
**Replies:** 9\
**Last updated:** [April 29, 2025, 2:04pm UTC](https://discuss.vllm.ai/t/grammar-cpu-bound-performance/372 "2025-04-29T14:04:47Z")

</div>

TL,DR: 100% cpu in top and GPU usage is very low for extended periods when using grammar. I’d expect to see perhaps 13\*100% on a 13 (real) core VM if it was perfectly efficient. Certainly more than one core. Are there …

---

## [Is structured output compatible with automatic prefix caching?](https://discuss.vllm.ai/t/is-structured-output-compatible-with-automatic-prefix-caching/456)

<div class="topic-metadata">

**Author:** [@g-eoj](https://discuss.vllm.ai/u/g-eoj)\
**Replies:** 1\
**Last updated:** [April 14, 2025, 6:56pm UTC](https://discuss.vllm.ai/t/is-structured-output-compatible-with-automatic-prefix-caching/456 "2025-04-14T18:56:53Z")

</div>

Imagine a scenario where there the prompt tokens are always the same but the extra\_body={"guided\_regex": argument can potentially change with a new request.

---

## [Do we have regression tests for structured output? Especially speed regression?](https://discuss.vllm.ai/t/do-we-have-regression-tests-for-structured-output-especially-speed-regression/339)

<div class="topic-metadata">

**Author:** [@lk-chen](https://discuss.vllm.ai/u/lk-chen)\
**Replies:** 0\
**Last updated:** [March 31, 2025, 11:58pm UTC](https://discuss.vllm.ai/t/do-we-have-regression-tests-for-structured-output-especially-speed-regression/339 "2025-03-31T23:58:00Z")

</div>

Did search in slack and found that “regression” happened several times before Correctness regression tests were not added until \[RFC\]: CI for key features · Issue #413 · vllm-project/vllm-ascend · GitHub. Do we plan…