Where to place CuTeDSL kernel

[luongthecong123]

Hi,
Can I submit the CuteDSL kernel in the same manner as the Triton one ?
Thanks,
Cong

[Weili-0234]

I have the same question as well🥹

[duoan]

same question?

[luongthecong123]

Hi folks, it seems that the submission is a pretty robust system.
You can add to the scripts/run_modal.py

image = (
    modal.Image.from_registry(f"nvidia/cuda:{tag}", add_python="3.12")
    .entrypoint([])
    .uv_pip_install(("flashinfer-bench", "torch", "triton", "numpy", "nvidia-cutlass-dsl")
...

Submit this under triton code and it runs fine, this works fine for any DSLs, pytorch works here too. Although I would argue that the flashinfer-bench library already installed cutedsl under the hood for us.

[duoan]

I put under python, it works as well just performance not good, just considering if that relates any missing libs

[Ubospica]

Please just put it under Python language. It should work. If there's any speed problem feel free to raise it @duoan .