ik_llama.cpp adds Gemma 4 graph parallelism

This looks like a real infrastructure unlock rather than a cosmetic tweak: if graph parallelism holds up, it could make 31B-class local inference practical for more people. The caveat is that the current signals are mixed, so speed gains need to be weighed against unstable outputs and quant-specific regressions.

• –The PR explicitly scopes the work to the dense Gemma4-31B model for now, and the maintainer says it already shows "very significant performance gains" versus layer mode.
• –Early comments in the thread suggest graph mode is exactly what people need for usable multi-GPU throughput on bigger Gemma 4 variants.
• –The PPL complaints around Unsloth quants look broader than just this fork; the repo itself warns that some Unsloth `_XL` GGUFs are likely to misbehave, so the bad perplexity may be partly a model-packaging issue rather than pure runtime failure.
• –Community reports also point to crashes, gibberish, or NaNs in graph mode on some setups, so this is promising but not yet a clean drop-in win.
• –If those edge cases get fixed, ik_llama.cpp could become one of the better local inference paths for high-end open Gemma 4 workloads.