libstrix update: one-command Docker image for Strix Halo, Qwen3.8-Flash-Next now runs at 40+ tok/s decode

Original title: libstrix update: one-command Docker image for Strix Halo, and Qwen3.8-Flash-Next now runs (1.6k prefill, 40+ tok/s decode)

The libstrix author released a one-command Docker image (ghcr.io/asf0/strix-vllm:latest) bundling a vLLM fork, the libstrix plugin and ROCm 10.1, so Strix Halo (gfx1151) users only need the amdgpu driver on the host. It adds Qwen3.8-Flash-Next support: its 102 GB n-gram (engram) table doesn't fit in 128 GB alongside the weights, so the fork streams it from disk and prefetches rows into the page cache; libstrix lifts prefill from ~827 to ~1,660 tok/s versus stock vLLM, and with MTP 2 reaches 1,539–1,617 tok/s prefill and 40–43 tok/s decode on 5.6k–100k prompts, with KLD 0.008–0.019 and GSM8K 93 vs 92. The image also ships recipes for Qwen3.6-35B-A3B, Qwen3.8-27B, LFM2.5-24B-A2B and gemma-4 26B/31B with a memory-overhead table per recipe, plus a heads-up that int8 decode copies are now limi…

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.