RuntimeError: vk::PhysicalDevice::createDeviceUnique: ErrorExtensionNotPresent

[smoky turtle]

RuntimeError: vk::PhysicalDevice::createDeviceUnique: ErrorExtensionNotPresent

[sage oriole]

Where do you get this

How

[bright sapphire]

Not every pod is vulkan capable

They dont test for it

If this is amd its certain not to have it, AMD is not good at supporting their cards so their MI300x has no vulkan driver

If its nvidia it depends on the docker container stuff the host has installed.

Btw @frail rivet if you wish to know if a host has vulkan to fix this issue across the nvidia fleet use https://koboldai.org/runpodcpp and in the KCPP_ARGS replace --usecublas with --usevulkan if the template works that host is good.

[smoky turtle]

Hi, my pod is running 1 x RTX 4090
14 vCPU 50 GB RAM
runpod/pytorch:2.4.0-py3.11-cuda12.4.1-devel-ubuntu22.04.
And I need to run GPU for rendering my simulation platform, the system returned error:
[2025-07-07 02:27:22.747] [svulkan2] [error] Vulkan is incompatible with your driver. You may not use the renderer to render, however, CPU resources will be still available.
scripts/bridge.sh: line 36: 609 Segmentation fault (core dumped)

[sage oriole]

I see, Look at Henky's answer

[smoky turtle]

sorry how should I check this exactly
"https://koboldai.org/runpodcpp and in the KCPP_ARGS replace --usecublas with --usevulkan if the template works that host is good." ?

The link brings me to the GPU selection page

[sage oriole]

With Henky's template right?

[smoky turtle]

Yes

[sage oriole]

Yeah just run the template I guess

With a pod of your gpu

Selection

[smoky turtle]

after this I should see KCPP_ARGS right?

[sage oriole]

Edit template then edit that args

On env variable

[smoky turtle]

Okay

[sage oriole]

Before deploying

[smoky turtle]

Oh I see

[smoky turtle]

I've created this template, but I just got disconnected so many times

I never experienced this on other templates

[bright sapphire]

My answer wasnt for him, it was so that runpod can more easily fix it

Its very hostbound

Some runpod hosts have vulkan some dont

[sage oriole]

ah

ya it'll be a pain for users to test, but ig thats the only way? unless.. contacting support to let them find you a pod

[bright sapphire]

Basically someone like dj has to use my template to detect which hosts dont have it and fix it

[frail rivet]

Working with our infra team to see what we can do to install Vulkan on every device that supports it

It would be part of a bigger rollout for some infra stuff, so I can't guarantee a date or timeline

[bright sapphire]

Vulkan itself is likely installed, its specifically the nvidia docker support that also passes trough the vulkan bins

[smoky turtle]

does the timeout indicate that pod isn't good for vulkan?

[bright sapphire]

If you mean that with using my template then probably, mine is only meant for running KoboldCpp not for whatever you want to run. I merely posted it so that runpods team has an easy way of testing vulkan.

[frail rivet]

We intend on testing Vulkan by just installing the SDK - https://docs.vulkan.org/guide/latest/checking_for_support.html#_ways_of_checking_for_vulkan

[bright sapphire]

The nvidia icd stuff has to be installed and thats auto injected

My template I ship the genetic mesa icd's but none of the runpod gpu's support that, they need the nvidia driver to add its own