nvcc fatal unsupported gpu architecture 'compute_86'

[szhengac]

When I install deepspeed from source, it gives me nvcc fatal unsupported gpu architecture 'compute_86'. My cuda version is 11.0. I can alleviate the issue by skipping 'compute_86' in

DeepSpeed/op_builder/builder.py

Line 265 in 6380ee3

args.append(f'-gencode=arch=compute_{num},code=sm_{num}')

. I think the current deepspeed installation may not handle cuda 11.0 well.

[sxjscience]

I'm also having this issue. The root cause should be that compute_86 is not supported in CUDA 11.0 and will only be available in CUDA>=11.1. Thus, we should be able to solve it by revising the logic to include the special treatment of CUDA 11.0

DeepSpeed/op_builder/builder.py

Lines 37 to 39 in 44bd538

	if torch.utils.cpp_extension.CUDA_HOME is not None and installed_cuda_version(
	)[0] >= 11:
	compute_caps += ";8.0;8.6"

[YaoBeiji]

nvcc fatal : Unsupported gpu architecture 'compute_86'，I have this issue too，but my CUDA version is 10.2

[loadams]

Closing this issue since this was a case where the CUDA version had to be updated to get support for newer devices, as is often the case.