Running H100 GPUs

Dear All-

I was trying to run PyFR on H100 GPU and It throws following error but I was able to run it with out any problem on 3090 the ampere architecture. I will appreciate if you have any suggestions.

Thanks!

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/base/backend.py”, line 207, in kernel

kern = kern_meth(\*args, \*\*kwargs)

       ^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/cuda/gimmik.py”, line 75, in mul

kern = self.\_build_kernel(kname, src, 'PP')

       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/cache.py”, line 44, in newmeth

res = cache\[key\] = meth(self, \*args, \*\*kwargs)

                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/cuda/provider.py”, line 41, in _build_kernel

mod = CUDACompilerModule(self.backend, src)

      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/cuda/compiler.py”, line 145, in _init_

cucode = backend.compiler.build('kernel', src, flags)

         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/cuda/compiler.py”, line 120, in build

code = self.nvrtc.compile(name, src, flags)

       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

File “/oscar/data/gk/kshukla1/NeuroNSE_DATA_JHTDB/PYFR/pyfr-venv/lib/python3.11/site-packages/pyfr/backends/cuda/compiler.py”, line 82, in compile

raise RuntimeError(log.value.decode())

RuntimeError: nvrtc: error: invalid value for --gpu-architecture (-arch)

Can you please confirm your CUDA version? It is likely your installation has an out-of-date version of nvrtc which is causing this issue.

Regards, Freddie.

Thanks @fdw It is cuda12.6

Can you confirm that nvrtc is from the exact same version?

Regards, Freddie.