search

McCloudS / subgen

Autogenerate subtitles using OpenAI Whisper Model via Jellyfin, Plex, Emby, Tautulli, or Bazarr
MIT License
453 stars 45 forks source link

Cuda error： no kernel image is available for execution on the device #82

Closed xhzhu0628 closed 3 months ago

xhzhu0628 commented 3 months ago

recently I've been running CUDA programs in Docker on Unraid. I have successfully installed the NVIDIA driver in Unraid and confirmed the GPU version and CUDA version (12.4) using the nvidia-smi command within Docker.

I'm using NVIDIA QUADRO K2200, CUDA 12.4

However, I'm still encountering an error when running the program: "no kernel image is available for execution on the device."

What could be the possible reasons for this, and how can I resolve it?

McCloudS commented 3 months ago

How are you set up? The docker container shouldn't let you run 12.4. Subgen (which uses: Whisper, CTranslate2, and Stable-ts) won't work on 12.4, which is why you get that issue. I am also running unraid.

nvidia-smi on the current subgen docker shows (just re-pulled it this morning): image

xhzhu0628 commented 3 months ago
屏幕截图 2024-04-14 221423 屏幕截图 2

I just installed the NVIDIA Driver plugin yesterday. Here are screenshots of the version in the Unraid plugin and the nvidia-smi output in Docker.

xhzhu0628 commented 3 months ago

How are you set up? The docker container shouldn't let you run 12.4. Subgen (which uses: Whisper, CTranslate2, and Stable-ts) won't work on 12.4, which is why you get that issue. I am also running unraid.

nvidia-smi on the current subgen docker shows (just re-pulled it this morning): image

it seems that I should use v545.29.06 instead

McCloudS commented 3 months ago

Yeah, try downgrading versions. Odd that my unraid doesn’t show your version, it shows me on the newest.

xhzhu0628 commented 3 months ago

image I downgraded the NVIDIA driver version, but the same error still occurs. Maybe related to cuda compute capability？ I have no idea about it.

McCloudS commented 3 months ago

Per https://opennmt.net/CTranslate2/hardware_support.html: “NVIDIA GPUs with a Compute Capability greater or equal to 3.5” for CTranslate2 but the K2200 has 5.0. I’m not sure…

xhzhu0628 commented 3 months ago

Ah, I think you’re right. Per https://opennmt.net/CTranslate2/hardware_support.html: “NVIDIA GPUs with a Compute Capability greater or equal to 3.5” for CTranslate2 and the K2200 has 3.0. I don’t think this will work for you.

Sorry.

https://developer.nvidia.com/cuda-gpus According to datasheet nvidia given, K2200 has 5.0 Compute Capability. But it's weird compared to other similar-level cards... I will change it and try again...

McCloudS commented 3 months ago

Yeah, I saw one nvidia document that said 3.0, but their official cuda page says 5. Looks like this is a similar issue to https://github.com/ollama/ollama/issues/1756

I’m still reading it.

McCloudS commented 3 months ago

I’m finding conflicting stuff showing that compute 5.0/Maxwell was kind of left behind. CTranslate2 doesn’t appear to build it for cuda12.

McCloudS commented 3 months ago

Asked around and see: https://github.com/OpenNMT/CTranslate2/issues/1666#issuecomment-2058390754

Compute 5.0 doesn't support CUDA 12, which CTranslate2 uses.

Sorry.