Ok, I finally made it:
❯ xrt-smi --version
Version : 2.25.0
Branch : HEAD
Hash : 3805b844aa6d2710763d22bd609ef54daf393415
Hash Date : Wed, 27 May 2026 08:29:44 -0700
amdxdna Version : 2.25.0_20260601, fdc88f9d166e3f43dc306564c09639c42f1bb9fb
virtio-pci Version : 7.0.10-201.fc44.x86_64
NPU Firmware Version : 1.1.2.64
❯ lemonade status
Server is running on port 13305
Property Value
--------------------------------------------------
Version 10.6.0
WebSocket Port 9000
Max Models/Type 2
Model Type Device Recipe Checkpoint
----------------------------------------------------------------------------------------------------
deepseek-r1-0528-8b-FLM llm npu flm deepseek-r1-0528:8b
It took quite some time to do it right, and FastFlowLM still needs to be installed manually (no RPMs), but the whole stack is running flawlessly now.
As far as I can remember, these are the main steps:
# install the dependencies needed to build XRT and FastFlowLM
sudo dnf install gcc \
gcc-c++ \
cmake \
boost-devel \
ncurses-devel \
doxygen \
python3-sphinx \
kernel-doc \
openssl-devel \
rapidjson-devel \
python3-devel \
pybind11-devel \
uuid-devel \
libuuid-devel \
systemtap-sdt-devel \
binutils \
binutils-gold \
binutils-devel \
libstdc++-static \
glibc-static \
rpm-build
# if you have OpenCL-ICD-Loader installed, you need to swap it for ocl-icd
sudo dnf swap OpenCL-ICD-Loader ocl-icd
sudo dnf install ocl-icd-devel
# clone XRT and its plugin repository
git clone --recursive https://github.com/amd/xdna-driver.git
cd xdna-driver
# they both depend on cmake3, which is really not needed
# change the line
# CMAKE=cmake3
# to
# CMAKE=cmake
# in build/build.sh and xrt/build/build.sh
# build XRT
cd xrt/build
./build.sh -npu -opt
# if all goes well, RPMs will be in the Release/ subdir
# install xrt...-base , xrt...-base-devel and xrt...-npu
# you will probably need to adjust your PATH to point to /opt/xilinx/xrt/bin and
# LD_LIBRARY_PATH to point to /opt/xilinx/xrt/lib
# now build and install the userspace plugin
cd ../../build
./build.sh -release -j(nproc)
# as above, the RPM will be in the Release subdir
Check if all went well:
xrt-smi --version
sudo xrt-smi examine
Now you need to clone FastFlowLM git repo, build it and install it in /opt:
git clone --recursive https://github.com/FastFlowLM/FastFlowLM
cmake --preset linux-default
cmake --build src/build -j(nproc)
sudo cmake --install src/build
Install Lemonade from its official page and start the service
sudo systemctl start lemond
lemonade backends # should show 'flm:npu' as installed
lemonade list # shows all available models; look for the ones ending in -FLM
lemonade pull <model>
lemonade run <model>
I most likely forgot some details; if you get stuck, post here and I’ll do my best to help. Good luck!
Thanks to @ciddan and @nadims for the dependencies info.