Blackwell Architecture support?

[iworry-dont]

Just checking if FlashInfer supports sm_120 (Blackwell) architecture now (NVIDIA GeForce RTX 5090 specifically). Your docs claim that FlashInfer only supports up to sm_90.

Thanks.

[yzh119]

Yes we added lots of blackwell kernels these months, but they are targeting B200 (sm_100a) instead of sm_120, both of them are called blackwell but they are different architectures, sm_100a kernels are not compatible for sm_120.

For sm_120, we will fallback to sm_80 kernels, which should still be functional.

Your docs claim that FlashInfer only supports up to sm_90.

Where did you see this, we should fix it.

[iworry-dont]

Hi - thanks for answering so quickly!

I saw the stuff about "up to sm_90" here: Installation

[yzh119]

Hmm that page is terribly out-dated, we should revamp :)

[iworry-dont]

For sm_120, we will fallback to sm_80 kernels, which should still be functional.

Wait - I built my PyTorch specifically and only for "120+PTX". That should cover backwards-compatibility for the sm80 architecture? Or no? Would I have to re-build PyTorch from source just to add sm80 kernel(s)? I'm self-taught here...

[yzh119]

No you don't have to re-build torch.

we will fallback to sm_80 kernels,

This means the kernel is written with sm_80 features, it is compatible with all later archs, when you JIT compile them, they will be compiled to target architectures.