Qwen fails to return valid structured output

LLM apps2 posts from 2 people2d active+00 posts in the last 7 days, 0 the 7 days before (fading)Unverified

Posts per day2 posts, Aug 31 to Oct 21

The posts behind it

2, newest first

GitHub issuesOct 7 [Bug]: Spec decode with structured outputs biases the sampling distribution (grammar-invalid drafts resampled from p instead of residual)### Your current environment The output of python collect_env.py (CPU flags, vulnerability list and GPU topology trimmed) ### 🐛 Describe the bug With structured outputs (a regex or other grammar) and draft-model speculative decoding with `draft_sample_method="probabilistic"`,...vllm-project/vllm · simon-veitner-redhat1 reactionsOpen ↗GitHub issuesAug 31 [Feature]: Upgrade XGrammar to >=0.2.4 and expose max_whitespace_cnt### 🚀 The feature, motivation and pitch vLLM currently declares the following XGrammar dependency: The generated test lockfiles currently resolve to: - CPU/CUDA/XPU: `xgrammar==0.2.3` - ROCm: `xgrammar==0.2.1` XGrammar added `max_whitespace_cnt` support for structural tags in...vllm-project/vllm · wangxuw3 reactions, 2 repliesOpen ↗

Companies and products named

About this problem

Evidence

2 posts from 2 people in 1 place, about 1 a week over 38 days. Mostly on vllm-project/vllm. Tools named alongside: vLLM, NVIDIA, GitHub.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts that state a pain and match the "structured-output" rule about Qwen. First post Aug 31, 2026, latest Oct 7, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

The same problem elsewhere