gpo.zugaina.org

Search Portage & Overlays:

sci-ml/openflowlm

Community fork of FastFlowLM with open NPU kernels for AMD Ryzen AI (XDNA2)

Screenshots

  • openflowlm-9999

    View      Download      Browse     License: MIT Apache-2.0 FastFlowLM-Terms   
    Overlay: stuff

ChangeLog

commit 37a4c5911626187a103b4ba8dfde9dbaaf3c33cd
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Oct 6 13:08:54 2026 +0200

sci-ml/openflowlm: new package, add 9999

OpenFlowLM is a community fork of FastFlowLM that replaces its closed
NPU kernels with open AIE designs built from source. It has no release
yet, so this is a live ebuild. Upstream's release process commits the
prebuilt open kernels to the branch each tag is cut from, so a
versioned ebuild can build without network or NPU access; one should
follow the first tag.

The open LLM kernels are compiled at build time in a uv venv holding
the mlir-aie and Peano wheels, which is why the build needs the
network. The BERT embedding sets are skipped: their export allocates
buffers on the NPU itself, and the build should not depend on the
device. granite42-3b and phi4-mini-4b are skipped because their
kernels fail to build at upstream HEAD.

The /usr/bin/oflm wrapper and env.d entry set OFLM_CONFIG_PATH
explicitly. Without it, oflm honours the FLM_CONFIG_PATH that
fastflowlm's env.d exports and reads FastFlowLM's model list.