[QST]Why does the 77_blackwell_fmha example not work on Blackwell architecture RTX5090 GPUs?

[luhan-wang]

Hi, cutlass team!

I'm trying example 77_blackwell_fmha which works fine on B200 GPUs but not on RTX5090 GPUs which are also Blackwell architecture, my compilation instructions are as follows:

nvcc -arch=sm_120a -I./include -I./tools/util/include -I./examples/77_blackwell_fmha -std=c++17 --expt-relaxed-constexpr \ ./examples/77_blackwell_fmha/77_blackwell_fmha_bwd.cu -o blackwell_fmha_bwd

The error is reported as follows:

###### B 16 H 16 Q 1024 K 1024 D 128 Backward Full #SM 170 [ ERROR: CUDA Runtime ] ./include/cutlass/cluster_launch.hpp:248: invalid argument Failed to launch the CUTLASS kernel. Last CUDA error is: no error [FAIL] tma : 0 TFLOPS/s

I would like to know why the above error is reported? Is 77_blackwell_fmha theoretically able to run on RTX5090?
Or is it only optimized specifically for the B200.

Sincerely thank you!

[tridao]

My understanding is that in cutlass "blackwell" = "datacenter blackwell" (e.g. B200), and "blackwell_geforce" = "consumer blackwell" (e.g. 5090). Those are 2 different archs

[thakkarV]

I would like to know why the above error is reported? Is 77_blackwell_fmha theoretically able to run on RTX5090?
Or is it only optimized specifically for the B200.

That's correct, example 77 is specifically for TCGen05 supported MMA architectures. Please use FA2 or example 41 for SM80/86/89/120

[luhan-wang]

My understanding is that in cutlass "blackwell" = "datacenter blackwell" (e.g. B200), and "blackwell_geforce" = "consumer blackwell" (e.g. 5090). Those are 2 different archs

Thank you very much for your reply, it's easy to understand!

[luhan-wang]

I would like to know why the above error is reported? Is 77_blackwell_fmha theoretically able to run on RTX5090?
Or is it only optimized specifically for the B200.

That's correct, example 77 is specifically for TCGen05 supported MMA architectures. Please use FA2 or example 41 for SM80/86/89/120

Thanks for the reply it was very helpful, I will try example 41 on the RTX5090