XGrammar

Structured-generation engine (Dong, Chen) adopted as the constrained-decoding backend in vLLM, SGLang, and TensorRT-LLM; XGrammar-2 (2026) extends to dynamic agentic workloads. Sibling Catalyst infrastructure: MLC-LLM universal deployment (23K★) and the PithTrain compact MoE trainer on FLAME compute.