Don't detect local GPU if `$DS_IGNORE_CUDA_DETECTION` is set (#7896)
Make this consistent over all OPs:
For cross-compilation we should not check the local GPU version.
This is also important when having multiple different GPUs in the
system, so checking only GPU 0 could be wrong.
It looks like checks for the version of CUDA should be moved out of this
and checks for compute capability vs CUDA version moved to `filter_ccs`
instead.
`CUDAOpBuilder.is_compatible` could call `compute_capability_args` to
see if the CCs after filtering turn up empty.
Co-authored-by: Olatunji Ruwase <tunji.ruwase@snowflake.com>
Co-authored-by: Masahiro Tanaka <81312776+tohtana@users.noreply.github.com>
DeepSpeed
26c954ff - Don't detect local GPU if `$DS_IGNORE_CUDA_DETECTION` is set (#7896)
