Setting up Jellyfin with Hardware Transcoding (NVIDIA GTX660)

[hutiwephy]

The problems:
- wrong driver version
installed 515
needed 470
- jellyfin won't start
Error: failed to start container "jellyfin": Error response from daemon: failed to create shim task: OCI runtime create failed: runc create failed: unable to start container process: error during container init: error running hook #0: error running hook: exit status 1, stdout: , stderr: Auto-detected mode as 'legacy' nvidia-container-cli: initialization error: nvml error: driver not loaded: unknown
- the menu for GPU allocation in jellyfin only allows 0 nvidia gpus independently of it being reserved or not in System -> Advanced -> GPU Allocation

The hardware:
Motherboard: Gigabyte GA-EX58-UD5-rev-10
CPU: i7-920
RAM: 16GB
Disks:
- WD3200BEKT (boot)
- WD5000AAKS (mainpool)
- WD5000AAKS (mainpool)
- WD5000AACS (mainpool)
GPUs:
- NVIDIA 9800GT (gigabyte)
- NVIDIA GTX660 (msi)

A dedicated guide for installing jellyfin on TrueNAS that actually worked would be apreciated (specially with nvidia transcoding on old cards)

help!
and
thanks!

 

[Gnome]

The problems:
- wrong driver version
installed 515
needed 470
- jellyfin won't start
Error: failed to start container "jellyfin": Error response from daemon: failed to create shim task: OCI runtime create failed: runc create failed: unable to start container process: error during container init: error running hook #0: error running hook: exit status 1, stdout: , stderr: Auto-detected mode as 'legacy' nvidia-container-cli: initialization error: nvml error: driver not loaded: unknown
- the menu for GPU allocation in jellyfin only allows 0 nvidia gpus independently of it being reserved or not in System -> Advanced -> GPU Allocation

The hardware:
Motherboard: Gigabyte GA-EX58-UD5-rev-10
CPU: i7-920
RAM: 16GB
Disks:
- WD3200BEKT (boot)
- WD5000AAKS (mainpool)
- WD5000AAKS (mainpool)
- WD5000AACS (mainpool)
GPUs:
- NVIDIA 9800GT (gigabyte)
- NVIDIA GTX660 (msi)

A dedicated guide for installing jellyfin on TrueNAS that actually worked would be apreciated (specially with nvidia transcoding on old cards)

help!
and
thanks!

From what I can tell, NVENC support was dropped for you card by nVidia on newer driver releases.
I base this on your card being missing from the support matrix for NVENC.
Only NVENC is supported by Jellyin

I get the impression from your two stated problems

The problems:
- wrong driver version
installed 515
needed 470

That you may already be aware of the above. So is your question along the lines of how you can install the older driver version?

 

[hutiwephy]

From what I've been seeing I can't do the downgrade so what's the alternative?

Drop TrueNAS and make my own with Debian that's the level of setting up that I wanted to avoid and While the CPU might take containers "Easily" VMs are a completely different story

Is it possible for me to swap the driver inside the container only (yes, I'm aware APT is gone)

If there was a way to switch or load multiple drivers that would be nice because the GT9800 is going to still trigger the jellyfin loading issue without version 390 loaded

Also with driver version 470 the NVENC encoder was supported by OBS on windows it's weird that jellyfin doesn't

 

[Gnome]

From what I've been seeing I can't do the downgrade so what's the alternative?

I unfortunately can't help you beyond what I've said so far with regards to options to downgrade.

If it were me, I'd first try creating a custom docker first (or looking for one that already runs the version you need).
I'm not an expert on docker containers, so it is unclear to me if Truenas determines the driver version in that situation.

Next up I'd try a VM with isolation and do your totally own thing. I suspect to do isolation and passthrough your motherboard would have to support VT-d or similar CPU instructions (like 99% sure but haven't done the due dilligence)

--

Lastly, I'd look on eBay and the like for alternatives, I know, I know, you might not want to spend the money or have other reasons, but the support is only going to get worse as time goes on.

I personally bought a Tesla P4 (data centre GPU) for very cheap on eBay ($60) which is basically a low power, low profile 1080 GTX with 6GB ECC RAM. But I've found that eBay is quite a treasure trove right now for older GPUs and Pascal and better all have NVENC (except maybe the absolute lowest end).

GPUs like the Tesla are/were terrible for bitcoin mining and well most things other than encoding really (since the GPU is locked to 75watt TDP), so they are especially cheap but there are Quadro cards that are equally suitable.

 

[hutiwephy]

The i7-920 and the MB support VT-d and it is enabled!

I definitely I'm not gonna spend a single penny on hardware because of broken software.

I'll make sure to post here if I find something for future reference and to avoid equal suffering for people in the same situation

So far the best option is a custom docker container (I have absolutely no idea how it works and never messed with it and I don't have the time to do it either the purpose of this test machine was to try and make it as simple as possible and to not have to mess with the OS or docker or kubernetes)

The VM option is most likely the only one that will work... Thanks TrueNAS!

 

[hutiwephy]

I said I would post if I had found a solution.

well that's impossible even CPU trans coding won't work because the GPUs prevent the container from starting and even though both motherboard and CPU support virtualization PCIe isolation is impossible making the VM a non viable solution

the only solution is for either TrueNAS allowing driver version selection NVIDIA providing universal drivers or making my own linux install

 

[winnielinnie]

From what I can tell, NVENC support was dropped for you card by nVidia on newer driver releases.

Does Nvidia actually do that? Removing features with newer drivers? I'm not a fan of the company itself, but even that seems too ridiculous for them.

 

[hutiwephy]

It's worse it removes the complete support for old cards resulting in Linux just ignoring the cards 3D capabilities and making it use the basic VESA capabilities such as text mode or some 2D rendering modes.

In order to drop support for features on old cards they mess with the libraries version forcing developers to use the newer version that doesn't work with the old drivers