vLLM adds target-token scoring support for DeepSeek-V4.1
October 21, 2026
The vLLM project has merged a pull request to preserve target-token scoring rows during DeepSeek-V4.1 replay. This improves the accuracy of inference-time scoring for the DeepSeek architecture.
HOW THIS AFFECTS YOU
●
builderYou can achieve more reliable scoring when running DeepSeek-V4.1 inference via vLLM.
●
researcherThis ensures better alignment between model scoring and replay accuracy.