GitHub ↗
← all posts
weeklystrix-haloryzen-ai-maxlinuxnpuhrxbosgame

This week in the engine

For two weeks the 1bit engine, the local LLM inference engine for the AMD Ryzen AI Max+ 395 (Strix Halo) on Linux, fixed open issues before building anything new: from 30 open engine issues to 6. Along the way a crash in the MTP drafter was found and fixed, the CI gained checks that catch whole classes of mistakes before review, and the Bosgame M5's dead USB-C ports turned out to have a one-line fix.

Why fix every issue first?#

The release is held until every issue is fixed. So from 31 October the work went to open issues first, oldest first, with new features after. Between 31 October and 15 November, 1,220 pull requests were merged into the engine, 1,764 in all since the repository started. Each one went through review and the merge loop, and every pin change also went through the hardware gate on a Strix Halo board.

What broke, and what fixed it?#

The MTP drafter crashed on the first request to any GGUF that carries its own MTP head, on the GPU route. The crash was found by serving such a file on engine main, and the fix came in with the llama.cpp pin (#2433, fixed by #2431).

The Lemonade fork sync failed on a cherry-pick that had become empty, because upstream already had the change. An empty cherry-pick now counts as a skip, not a failed port (#2454).

What do the new CI gates check?#

Each of these answers one question about a branch before a reviewer has to:

  • pre-pr-checks runs the repository's own checks on a branch, including the stale-branch and registry-loss checks, and says when the tree is dirty (#2396, #2401).
  • branch_stale_check names the files a branch that is behind main would write back over newer work (#2398).
  • guard_loss_check reports a refusing or asserting condition that a change deleted (#2429).
  • tests_referenced now requires every test script to be invoked by a registered test, not just mentioned somewhere (#2460).

What moved on the GPU route?#

The GPU route runs the Radeon 8060S through HRX with our own HIP kernels (HRX). Its two pins moved forward, each through the hardware gate:

  • the gpu-kernels pin moved twice, once for the prefill route on the 27B Q4_K model and once for decode (#2449, #2453);
  • the llama.cpp pin now reads the recurrent state of gated delta-net models in place during decode, instead of copying it each step (#2461).

The NPU route, our own runtime for the XDNA 2 NPU, is described on NPU.

Do the Bosgame M5's USB-C ports work on Linux now?#

Yes. On the Bosgame M5 (board AXB35-02), three of the four USB controllers sleep in D3cold under Linux, so only the front USB-A port worked. A one-line udev rule keeps them awake, and with it every external port enumerated in a plug test on the hardware on 14 November (BIOS 1.10). USB4 devices are not tested yet. The rule, the check and the per-port table are in Bosgame M5 USB-C not working on Linux: the fix, and the full record is on AXB35 USB ports.

The previous update is This week in the engine, 30 October. The code is in the 1bit engine repository, and to run it from a USB stick, see Install.

Frequently asked questions#

What is the 1bit engine?#

A model-agnostic local LLM inference engine for the AMD Ryzen AI Max+ 395 (Strix Halo) on Linux. It serves GGUF models on the Radeon 8060S through HRX and models on the XDNA 2 NPU, as a backend of Lemonade Server.

Is the release out?#

No. The release is held until every open engine issue is fixed.

Does the engine use the AMD NPU on Linux?#

Yes. It has its own NPU runtime for the XDNA 2 NPU, next to the GPU route through HRX.