Nvidia 5080 Not being detected as CUDA GPU

[Dan64]

Hello Selur,

  I have the same problem. Hybrid is not detecting my new GPU RTX 5070 Ti.
  Also NVEnc is not able to recognize it, I opened an issue for this: https://github.com/rigaya/NVEnc/issues/675

  The RTX 50xx need CUDA 12.8, fortunately few days ago was released a pytorch build for windows with CUDA 12.8 support.
  The new pytorch build is using the CUDA library: 12.8.57 and the cuDNN library: 9.7.1.26.

  To be able to use the card with vs-deoldify I had to run the following commands:

.\python -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128

  Moreover in the last torch distribution VapoursynthR70_torch_2025.02.22.7z was missing the package  opencv-contrib so I had to install it

.\python -m pip install opencv-contrib-python

Given that my RTX 3060 has 3584  shader processors (at 1320 MHz) while the RTX 5070 Ti has 8960 processors (at 2295 MHz) I was expecting an increase in speed of at least 3X, but in my preliminary tests the increase in speed is only 2.2X. I had not time to optimize the GPU and mybe I will able to improve the speed further. But at the moment I'm busy to get all packages working with this new GPU.

Please provide me the link to the new Hybrid version supporting the RTX 50xx card, when it will be available.

Thanks,
Dan

[Selur]

try the dev

[Dan64]

It works!
I got the following log output:

Detected NVIDIA PureVideo compatible cards: NVIDIA GeForce RTX 5070 Ti
Detected vfwDecoders with 32bit:   vidc.cvid   vidc.i420   vidc.iyuv   vidc.mrle   vidc.msvc   vidc.uyvy   vidc.yuy2   vidc.yvu9   vidc.yvyu
Detected vfw64BitDecoders:   vidc.i420   vidc.iyuv   vidc.mrle   vidc.msvc   vidc.uyvy   vidc.yuy2   vidc.yvu9   vidc.yvyu
   Avisynth+ is available,..
    DGDecNV available,..
   Vapoursynth is available,..
    DGDecNV available,..
    VSGAN available,..
    vsDPIR available,..
    vsRIFE (torch) available,..
    vsBasicVSR++ available,..
    vsRealESRGAN available,..
    vsSwinIR available,..
    vsHINet available,..
    vsAnimeSR available,..
    vsFeMaSR available,..
    vsSCUNet available,..
    vsCodeFormer available,..
    vsGRLIR available,..
    vsDDColor available,..
    vsHAVC available,..
    vsProPainter available,..
    vsDeepDeinterlace available,..
    vsMFDin available,..
    VSMLRT available,..
    vsDPIR (mlrt) available,..
    vsRIFE (mlrt) available,..
    vsSAFA (mlrt) available,..
    vsSCUNet mlrt) available,..
    vsSwinIR (mlrt) available,..

Thanks,
Dan

P.S.
I hope that also rigaya will fix NVenc...

[Dan64]

I completed the migration of torch packages to support the RTX 50 series.

I installed:

tensorrt-10.9.0.34
tensorrt_cu12-10.9.0.34
tensorrt_cu12_libs-10.9.0.34
tensorrt_cu12_bindings-10.9.0.34
torch_tensorrt-2.7.0.dev20250316+cu128
torch-2.8.0.dev20250314+cu128
torchaudio-2.6.0.dev20250315+cu128
torchvision-0.22.0.dev20250315+cu128

Also is necessary to install Pytorch-Correlation-extension reported in: #1078

Tested:

• vsdeoldify
  
• vsbasicvsrpp
  
• vsanimesr (with TensorRT)
  
• vscodeformer
  
• vsrealesrgan (with TensorRT)
  
• vspropainter
  

Dan

[Selur]

Why torchaudio? Doesn't seem to be necessary at all.

[Dan64]

It was included in the pip command. 




but in Hybrid is not used. It can be removed from the installation command

.\python -m pip install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/cu128

Dan

[tailland]

(brain fart: deprecated cuda32 on the 50x0 being the reason for this)

[Dan64]

rigaya fixed the issue on RTX 50: #675

Dan

[Selur]

Uploaded a new dev which comes with the new NVEncC version.

[The-Architect]

✅ Issue Solved on RTX 5080 — Full Compatibility Achieved
Hey everyone, I was hitting the same CUDA sm_120 block and runtime errors on the RTX 5080. After deep debugging and patching attempts, I finally found a clean workaround:

1. Install PyTorch Nightly with CUDA 12.8:
   bash
   CopyEdit
   
   pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
2. Install or update
   
   bitsandbytes
   manually (for CUDA 128 support):
   • https://github.com/bitsandbytes-cuda/bit...s/releases
     
3. Fix the missing
   
   libbitsandbytes_cuda128.so
   :
   • Search for it (possibly in ComfyUI or Forge folder) and manually copy it to your target venv path:
     swift
     CopyEdit
     
     your_venv_path/lib/site-packages/bitsandbytes/
4. (If needed) Downgrade Triton:
   bash
   CopyEdit
   
   pip install triton==3.1.0

Everything now runs smoothly on my RTX 5080, including WebUI Forge and Diffusers. Let me know if anyone needs help replicating this.
Repo with fix and driver insights coming soon. Marking this as Solved.
— The Architect