Skip to content

Older NVIDIA GPUs get a CUDA torch build with no kernels for them #732

Closed

Description

@thcp

Split out of #723.

What goes wrong

On a GTX 970 (compute capability 5.2) with a driver reporting CUDA 13, first-run setup installs CUDA torch, verification fails with "no kernel image is available for execution on the device", and the app falls back to CPU. The GPU would work with an older torch build.

Why

For a CUDA 13 driver the wheel list is cu128 only. The cu128 torch build has no kernels below sm_61 (Pascal), so Maxwell cards such as the GTX 9xx series can never pass. The cu124 and cu118 builds are never tried.

Constraints

  • The compute capability is already detected before choosing a wheel.
  • Whether cu124 or cu118 still ship sm_50 and sm_52 kernels on Windows has to be checked against the published wheels, not assumed.

Activity

  1. added
    bugSomething isn't working
    on Sep 30, 2026
  2. self-assigned this
    on Oct 1, 2026
  3. added a commit that references this issue on Oct 1, 2026
    d78fa34
  4. added a commit that references this issue on Oct 2, 2026
    939d66e
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't working

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions