vllm
[Gemma4] Allow per-layer attention backend selection for heterogeneou…
#38891
Open
Open