Papers and posts.
What we have built, how we measured it, and what did not work. Every number links back to a method you can check.
2026
Paper · Inference systems
Flux-GPU: measured autotuning for local inference on consumer GPUs
A runtime that screens llama.cpp configurations, confirms winners with interleaved runs and Welch's t-test, and stores them per hardware fingerprint. +68.3% generation on an RX 580; 12B models on an 8 GB card.
Blog · Measurement
Measure, then tune: how we decide a change is real
Five rules every result on this site goes through, and the three mistakes they caught in a single week.
Blog · Inference systems
One queue, 68% faster: what an autotuner found on an RX 580
Not flash attention, not batch size, not a quantized cache. The winning change was which Vulkan queue the work went to.