vLLM adds CUDAGraphMode.FULL_DECODE_ONLY support for Voxtral Realtime
September 14, 2026
The vLLM inference engine now supports CUDAGraphMode.FULL_DECODE_ONLY for the Voxtral Realtime model. This integration enables optimized execution paths within the runtime for this specific model architecture.
HOW THIS AFFECTS YOU
●
builderYou can achieve better inference performance for Voxtral Realtime using this optimized mode.
●
researcherThis provides a signal that Voxtral Realtime is moving toward production-ready deployment.