Skip to content

CUDA error on RTX 5080 (sm_120/Blackwell) — PyTorch not compiled for sm_120 #396

Closed

Description

@fanhua2022-commits

VoiceBox v0.3.0, Windows 11, NVIDIA GeForce RTX 5080
The changelog for v0.3.0 mentions "CUDA support for RTX 50-series GPUs", but generation still fails with:
torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device

The server log shows the bundled PyTorch only supports up to sm_90:
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90.
NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation.

RTX 50-series (Blackwell) requires PyTorch compiled with CUDA 12.8+ and sm_120 support (torch 2.7.0+cu128 or newer). The bundled torch in the CUDA backend needs to be updated to include sm_120 kernels.
Workaround I found for other projects: manually installing torch-2.7.0+cu128-cp310-cp310-win_amd64.whl from the official PyTorch wheel index resolves the issue, but this is not possible with VoiceBox's bundled environment.
Please update the bundled PyTorch in the CUDA backend to a version compiled with sm_120 support.

Activity

  1. pritam4249 commented on Apr 12, 2026

    @pritam4249

    Yes, similar error while running in 5060ti, kindly fix this issue in next update. We need to use CPU currently.

  2. duquevazzaro commented on Apr 13, 2026

    @duquevazzaro

    I've compiled a version that includes CUDA support for the RTX 50 Series. I have an RTX 5080, and it works perfectly. If anyone's interested, I'll upload the release to my GitHub so you can try it out and give me some feedback.

  3. omnsnr commented on Apr 14, 2026

    @omnsnr

    I've compiled a version that includes CUDA support for the RTX 50 Series. I have an RTX 5080, and it works perfectly. If anyone's interested, I'll upload the release to my GitHub so you can try it out and give me some feedback.

    That would be great

  4. duquevazzaro commented on Apr 14, 2026

    @duquevazzaro

    I've compiled a version that includes CUDA support for the RTX 50 Series. I have an RTX 5080, and it works perfectly. If anyone's interested, I'll upload the release to my GitHub so you can try it out and give me some feedback.

    That would be great

    In addition to improved compatibility with all RTX 50 Series cards, I’ve implemented 10 additional improvements, including performance enhancements, multilingual app translation, faster load times, and an integrated CUDA library, among others... It’s worth the wait. I’ll have it ready and compiled for Windows in a few hours. Stay tuned

  5. duquevazzaro commented on Apr 16, 2026

    @duquevazzaro

    VoiceBox v0.3.0, Windows 11, NVIDIA GeForce RTX 5080 The changelog for v0.3.0 mentions "CUDA support for RTX 50-series GPUs", but generation still fails with: torch.AcceleratorError: CUDA error: no kernel image is available for execution on the device

    The server log shows the bundled PyTorch only supports up to sm_90: The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_61 sm_70 sm_75 sm_80 sm_86 sm_90. NVIDIA GeForce RTX 5080 with CUDA capability sm_120 is not compatible with the current PyTorch installation.

    RTX 50-series (Blackwell) requires PyTorch compiled with CUDA 12.8+ and sm_120 support (torch 2.7.0+cu128 or newer). The bundled torch in the CUDA backend needs to be updated to include sm_120 kernels. Workaround I found for other projects: manually installing torch-2.7.0+cu128-cp310-cp310-win_amd64.whl from the official PyTorch wheel index resolves the issue, but this is not possible with VoiceBox's bundled environment. Please update the bundled PyTorch in the CUDA backend to a version compiled with sm_120 support.

    Ready guys https://github.com/duquevazzaro/clonvoice-releases/tree/main
    Buy me a coffee :)

  6. capy-ai-staging commented on Oct 7, 2026

    @capy-ai-staging
    Contributor

    This was fixed by #316 and #401 (thanks @mvanhorn): the Windows CUDA backend moved from cu126 to cu128 wheels with sm_120 in the build, so RTX 50-series (Blackwell) cards have worked since Voicebox 0.4.0. If you are still on 0.3.x, please update and re-download the CUDA backend from Settings. Closing as fixed; if you still hit "no kernel image is available" on 0.4.0 or later, please reopen with the log.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions