vLLM Fixes Gemma 4 Encoder Compatibility via max_soft_tokens
October 24, 2026
A new pull request in the vLLM repository resolves an encoder computation graph issue for Gemma 4 models. This fix involves correctly handling the max_soft_tokens parameter to ensure stable inference runtime support.
HOW THIS AFFECTS YOU
●
builderYou can now deploy Gemma 4 models using the vLLM inference engine without encoder errors.