#GPU on HAOS for voice

1 messages · Page 1 of 1 (latest)

vale forum
#

I have HAOS Generic x86-64 installed on a system with an nvidia 1660-ti. Does homeassistant support using the gpu for voice or do i need to install it via vm or docker?

frank bison
#

There are a few references to NVIDIA in the code: https://github.com/search?q=repo%3Ahome-assistant%2Fbuildroot nvidia&type=code
What I can tell you is that the OVA (VM) version does not seem to have any drivers for it included.

# find / -iname "*nvidia*"
/etc/apparmor.d/abstractions/nvidia
/etc/apparmor.d/abstractions/opencl-nvidia
/etc/apparmor.d/local/nvidia_modprobe
/etc/apparmor.d/nvidia_modprobe

# find /lib/modules/ -iname "*nvidia*"
#

# lsmod
Module                  Size  Used by
bluetooth             819200  0
ecdh_generic           16384  1 bluetooth
ecc                    40960  1 ecdh_generic
nft_chain_nat          12288  23
xfrm_user              57344  1
nft_compat             20480  58
nf_tables             356352  654 nft_compat,nft_chain_nat
cfg80211             1257472  0
sch_fq_codel           20480  2
virtio_balloon         32768  0
dmi_sysfs              16384  0
qemu_fw_cfg            16384  0
virtio_console         40960  1

Try those commands yourself on x86 and see what you get.
Some container have a libdrm_nouveau.so* file but if that is helpful I don't know

# find / -iname "*nouveau*"
/mnt/data/docker/overlay2/0eed640bd4a2b167600f63dbf9e478803bacdc59bcf551164adf9841d027f6fa/diff/usr/lib/libdrm_nouveau.so.2.123.0
/mnt/data/docker/overlay2/0eed640bd4a2b167600f63dbf9e478803bacdc59bcf551164adf9841d027f6fa/diff/usr/lib/libdrm_nouveau.so.2
/mnt/data/docker/overlay2/86a73f24ddba00017491829136ea3a0cf3ae656fdb49dcb92c603721e324a04a/merged/usr/lib/libdrm_nouveau.so.2.123.0
...
vale forum
#

the most I can find are posts for Frigate saying HAOS standalone does not support GPU. For Voice it seems most people either run a separate ollama server or run them both as vms or containers.

Seems like an oversight to develop voice and local ai without the hardware support. Though I suppose target hardware are mini PCs and Pis.

I'd expect support to come eventually but I couldnt find any mention.

leaden token
#

what part of voice are you trying to accelerate? if its whisper then a suggestion would be to run proxmox as os. run haos on 1 vm and a generic linux install (say ubuntun?) on another. pass through the gpu to the linux vm and run whisper with acceleration on that. then add that whisper to wyoming intrgration.
I run whisper on not only a separate VM but a separate server 🙂

frank bison
#

I'd actually recommend a CT for GPU related stuff whenever possible.

leaden token
frank bison
#

LXC container.

#

This way the node as well as other CTs can use the GPU too. Passing it through can be a massive pain.

leaden token
#

ah yeah an LXC container works too.

lucid fog
#

Docker can also pass /dev/ nodes (aka the gpu node) then nothing in the host but the kernel module would matter

#

But i dont know if Ha will allow you to run a container with the --priviledged tag you can always do that manually