Function g4_batch_on 

pub fn g4_batch_on() -> bool

Gemma-4’s prompt on the batched wgpu graph (CMF_G4_BATCH=0 keeps the chunked host prefill it used before).