Error when building flashinfer-jit-cache from source.

[pavank63]

Hi Team, we are trying to build the flashinfer-jit-cache from the source repository and the AOT compilation step currently fails with the following issue

ptxas /tmp/tmpxft_00018a8b_00000000-10_quantization.compute_120.ptx, line 15870; error   : Instruction 'cvt with .e2m1x2' not supported on .target 'sm_120'
ptxas fatal   : Ptx assembly aborted due to errors

I have FLASHINFER_CUDA_ARCH_LIST: "7.5 8.0 8.6 8.7 8.9 9.0a 10.0 10.0a 12.0" and the builds succeeds when i remove 10.0 and 12.0 from the list. Does that mean the sm_100 and sm_120 do not support .e2m1x2 ?

Are there any flags that I can use to skip just .e2m1x2 compilation and still do rest of AOT for sm_100 and sm_120 or should I just remove them from the FLASHINFER_CUDA_ARCH_LIST and jit compile for these two architectures ?

[EmilienM]

ping @yzh119

[yzh119]

Hi @pavank63 @EmilienM what's the cuda version you are using?

[pavank63]

We are using CUDA 12.8.1