[BUG] `88_hopper_fmha_fp8` example fails to compile on some CUDA archs

[Flamefire]

Which component has the problem?

CUTLASS C++

Bug Report

Describe the bug
I'm doing a fat build of CUTLASS for a cluster with different GPUs including a pre-7.5 one.

However the FP8 variant of the 88_hopper_fmha example fails to compile with many errors like

ptxas /tmp/tmpxft_0009b33b_00000000-7_88_hopper_fmha.compute_70.ptx, line 414521; error : Feature 'ex2.f16/f16x2' requires .target sm_75 or higher

Steps/Code to reproduce bug

Configure with cmake ... -DCUTLASS_NVCC_ARCHS="80;70"

Expected behavior

All examples build successfully to verify the build worked.

[hwu36]

This is a Hopper kernel. It needs to be compiled with 90a. You are using 80 and 70 which are not supported.

[Flamefire]

I understand that. But if you are building for 8.0 shouldn't the configure process ensure that this test/example isn't added?
Because now it is impossible to compile any example with <90a as the build will fail on this example.

[alihassanijr]

Sorry no, as far as I know there's not really anything that automatically prunes targets according to specified architectures.

Profiler and cutlass library are exceptions, they stamp out only compatible kernels, and those don't really get built as a fatbin in the same way.

Examples are expected to compile even on older unsupported arches without error, but with an empty kernel and throw a runtime error. It might be something that's missing in this example.

Can you confirm this is the only build error you're hitting, and if not, what other examples fail to build?

[alihassanijr]

@hwu36 this is due to missing CUTLASS_ARCH_MMA_SM90A_ENABLED guards in the kernel. I'll patch this and the distgemm one, since they're technically the same issue.

[Flamefire]

Can you confirm this is the only build error you're hitting, and if not, what other examples fail to build?

Tested 4.1.0 the other day and this and #2558 were the only examples that failed to compile

[alihassanijr]

#2567 should fix both. Would be great if you could apply the patch and test it.

[Flamefire]

I downloaded the whole branch and rebuild with -DCUTLASS_NVCC_ARCHS=60 without errors.

I also tested my previous configuration of 4.1.0 with that applied and it does succeed now