ROCm 5.5 "cumputeCapability" issue

[cornpo]

Hello,

With ROCm 5.5, 2.1.0.dev20230502+rocm5.4.2, Triton 2.0.0.post1, pytorch-triton-rocm 2.0.2

Running oobabooga text webui loading a quantized model.

python: /project/lib/Dialect/TritonGPU/Transforms/Combine.cpp:870: int {anonymous}::{anonymous}::computeCapabilityToMMAVersion(int): Assertionfalse && "computeCapability > 90 not supported"' failed.`

It works with cuda. It will load the model in Triton. But gives that error at infer time.

[psinger]

@cornpo running into the same, did you find a solution?

[cornpo]

Not for the Triton stuff. Cuda and old-cuda worked, as I remember. Haven't tried anything new since beginning of June.

Good luck.

[phaylali]

anyone found a solution for this?

[binarman]

Hi everyone!

Triton 2.0 is relatively old version of triton, is it mandatory to use it for you?