Bus use: stock llama.cpp reached ~145GB/s on Ornith. These engine tweaks reach ~171GB/s over a whole pass and 234GB/s in its best kernel. The memory floor for a pass is ~17.4ms, this implementation reaches 23.5ms so still some headroom left at the same quant.

Caveats: custom ggml + libllama under stock llama-server Speed moves ~2.5% run to run and more by prompt. Pooled routing changes the model, a preliminary and partial SWE-bench run showed about 11/19 resolved rate with my pi harness, 10m per task.

Draft head: only the 65,536 most common tokens of 248k, stored at Q4_0. Used for drafts only, every token is still verified by the full Q8_0 model so the output does not change. It only moves draft acceptance rate.

Next lever is MTP and Ornith ships its own draft block. It drafts 3 tokens, then one main pass verifies 4 positions. The trunk is read once per pass, not once per token. So, a thin trunk makes verification cheap. ~2.7 tokens kept per pass.

Doing some AMD Halo Strix testing... First is picking a base model that fits the strix's architecture and is able to run as optimally as the hardware allows. Ornith 1.5 35B A3B fits this description as it is small-trunk MoE, allowing for small experts on the shared memory bus.