RPM packaging for FastFlowLM on Fedora Linux.
FastFlowLM runs large language models on AMD Ryzen AI XDNA2 NPUs.
Add the COPR repo:
sudo dnf copr enable alessandrolattao/fastflowlmInstall the driver and runtime (DKMS builds the kernel module automatically):
sudo dnf install fastflowlmReboot to activate the NPU driver:
sudo rebootDownload the proprietary NPU kernel binaries (requires internet access to GitHub):
sudo flm-fetch-kernelsVerify the NPU is detected and working:
flm validateRun a model:
flm run llama3.2:1bEPEL is required for dkms. Enable it before adding the COPR repo:
sudo dnf install epel-release
sudo dnf install dkms kernel-develAdd the COPR repo:
sudo dnf copr enable alessandrolattao/fastflowlmInstall the driver and runtime (DKMS builds the kernel module automatically):
sudo dnf install fastflowlmReboot to activate the NPU driver:
sudo rebootDownload the proprietary NPU kernel binaries (requires internet access to GitHub):
sudo flm-fetch-kernelsVerify the NPU is detected and working:
flm validateRun a model:
flm run llama3.2:1b- AMD Ryzen AI CPU with XDNA2 NPU (Strix, Strix Halo, Kraken, Gorgon Point)
- Linux kernel >= 6.10. On kernel < 7 the
xdna-driver-dkmsmodule is pulled in automatically (it is the only driver available). On kernel 7+ the kernel's built-inamdxdnadriver is used by default; the DKMS module is optional there (see Newer NPUs on kernel 7+) - NPU firmware >= 1.1.0.0 (installed automatically by
xdna-driver) - Unlimited memlock limit: the NPU requires large DMA buffers locked in physical RAM to load model weights. The default kernel limit (64 KB) is far too low.
xdna-driversets this automatically via/etc/security/limits.d/99-amdxdna.conf. Log out and back in after install for it to take effect
On kernel 7+ the package uses the kernel's in-tree amdxdna driver by default.
That driver tracks the kernel and can lag behind AMD's: on a recent NPU (e.g.
NPU3) or with newer firmware, the in-tree driver may fail to detect the device.
In that case, install the out-of-tree driver module, which ships AMD's newer
amdxdna 0.15 and overrides the in-tree one:
sudo dnf install xdna-driver-dkms
sudo rebootTo revert to the in-tree driver, remove it:
sudo dnf remove xdna-driver-dkms
sudo rebootThis only affects the kernel module. It is not needed if flm validate
already detects your NPU.
Note: the
xdna-driverpackage version (e.g.2.25.0) is not the same as theamdxdnadriver version reported byflm validate(e.g.0.6or0.15). On kernel 7+, after a normaldnf update,flm validatekeeps showing the in-tree driver version (0.6/0.8), which is expected. It only switches to the newer out-of-tree version after you installxdna-driver-dkmsand reboot.
The NPU kernel binaries (.so + xclbin files) are proprietary and cannot be
redistributed in this package. flm-fetch-kernels downloads them from the
official FastFlowLM GitHub release and installs them to /opt/fastflowlm.
By running flm-fetch-kernels you accept the
FastFlowLM Proprietary Binary License v2.0.
| Package | Description |
|---|---|
fastflowlm |
FastFlowLM CLI and runtime (MIT license) |
xdna-driver |
AMD XDNA userspace driver, XRT libraries, NPU firmware, DKMS kernel module (Apache-2.0) |
This is an unofficial, community-maintained repository. It is not affiliated with, endorsed by, or supported by FastFlowLM Inc. or AMD.
- FastFlowLM software is developed and owned by FastFlowLM Inc.. This repo only provides packaging.
- The NPU kernel binaries downloaded by
flm-fetch-kernelsare subject to the FastFlowLM Proprietary Binary License v2.0. Read it before use. - Use at your own risk. No warranty, no support, no liability.
fastflowlm/ fastflowlm.spec + flm-fetch-kernels
xdna-driver/ xdna-driver.spec (XRT + firmware + DKMS kernel module)
.copr/ Makefile used by COPR to generate SRPMs
docs/ Build and publishing guides
Makefile Targets: srpm, copr, bump, clean