llama.cpp graph opts boost Gemma 4 prompt processing

llama.cpp has picked up graph optimizations for Google's Gemma 4 models that substantially raise prompt-processing speed on the E2B and E4B sizes.

Prompt processing is the stage where the runtime ingests and encodes the input before it starts generating tokens. For interactive chat and batch jobs that stage often dominates wall-clock time, so graph-level work that shortens it matters directly to anyone running these models on local hardware through the popular C++ engine.

The improvement was reported against the E2B and E4B variants specifically, with accompanying benchmark figures shared on the ggml-org project showing large throughput lifts in that phase. Community reaction on the repository was immediate and positive.

The result is faster time-to-first-token behavior for Gemma 4 users who already rely on llama.cpp for local and edge inference, without changing the models themselves.