Fixing Tool-Call Accuracy in vLLM

Tool call accuracy is crucial for agentic workloads in serving open-weight models, but many providers and inference engines struggle to achieve 100% accuracy. Constrained decoding, a technique added in vLLM 0.31, can guarantee schema-adherence and achieve 100% accuracy with newer Kimi and DeepSeek models. This is especially important for agentic workloads that require multi-turn sessions and rely on accurate tool calls to produce high-quality results.