Dose Triton supports new features of Blackwell for RTX5090 and 5080?

[jt-zhang]

I saw that Triton supports new features of Blackwell, but is it also compatible with RTX5090 and 5080?

[DJCool1]

Well, I followed the github instructions today and successfully compiled it in a venv but I'm still getting

'sm_120' is not a recognized processor for this target (ignoring processor)

When I try to use it but am using Torch nightly on WSL / Linux sucessfully otherwise on the RTX 5080 without sm_120 errors when not using Triton so it seems like the support is spotty?

[DJCool1]

Update: thanks to the discord server, @drisspg noted to remove PyTorch-triton and that fixed it for me so in answer to your question, yes it now works on a RTX 5080.

[jt-zhang]

@DJCool1 Thank you for your reply! However, what I mean about the new features of Blackwell is the FP4 micro-scaling Matmul. Have you successfully tried it on RTX 5080 using Triton? Thank you again.

[DJCool1]

Sorry, I don’t know; my work does not involve FP4 Matmul.