Engineering Records

Measurement-driven records of systems work: what was tried, what it measured, what shipped, and what didn't work — with the numbers.

Optimizing DeepSeek V4 Flash Inference on a Strix Halo Pair: An Engineering Record

A week-long ROCm optimization campaign on two Strix Halo nodes: MXFP4 routed-expert kernels, a bit-exact k-row speculative verifier, and a measured tour of what didn't work — 11.35 → 14.67 tok/s decode, 1.7× prefill.

August 14, 2026