Skip to content

SageAttention cu130 / Torch 2.10 wheel missing SM120 architecture #22

Description

@ironfisc-source

Environment: Windows 11, RTX 5070 Ti (compute capability 12.0), Python 3.13.11, PyTorch 2.10.0+cu130.

The current sageattention-2.2.0+cu130torch2.10-cp313-cp313-win_amd64.whl fails on Blackwell with cudaErrorNoKernelImageForDevice for longer attention workloads.

packages/sageattention.yml has global arch_list: "8.0 9.0". The cu130/Torch 2.9 and Torch 2.11 combinations override this with "8.0 9.0 10.0 12.0", but the cu130/Torch 2.10 combination does not.

Inspecting the installed _fused.cp313-win_amd64.pyd and _qattn_sm89.cp313-win_amd64.pyd shows only sm_80 and sm_90a markers, with no sm_120/sm_120a.

Minimal reproduction with GQA 32:8, head_dim=128, L=4096 produces CUDA error: no kernel image is available for execution on the device, while L=128 works.

Adding arch_list: "8.0 9.0 10.0 12.0" to the cu130/Torch 2.10 matrix entry and rebuilding the wheel appears to be the expected fix.

Activity

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

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions