vLLM Adds Stop Sampler Recompiles for DiffusionGemma
November 13, 2026
The vLLM repository has merged support for DiffusionGemma to prevent stop sampler recompiles during serving. This improvement helps stabilize inference performance for diffusion models.
HOW THIS AFFECTS YOU
●
builderThis reduces latency spikes when serving DiffusionGemma models.