I wrote a from-scratch Vulkan inference engine for one model (Qwen3.6-35B-A3B) on RDNA3 — 1.44x llama.cpp decode, token-exact parity

Article automatically generated from technical news.

TL;DR — I hand-wrote a Vulkan compute engine specialized for a single model (Qwen3.6-35B-A3B) on RDNA3. It decodes at 190.7 tok/s vs llama.cpp's 132.3 on the same GGUF and the same card — 1.44x — with token-for-token identical greedy output. Source: https://github.com/ryanmurf/qwen-kernel What it is Not a llama.cpp fork. It's a from-scratch Vulkan inference engine + serving stack that does exactly one model and does it fully specialized. Inspired by KernelBench Mega (which is