For some reason 4.7 UCX 1.7 is not built anymore with --enable-mt option and codes which work fine with MOFED 4.6 UCX (e.g. tensorflow) crashes immediately with MOFED 4.7. Rebuilding rpm with --enable-mt fixes the issue but it was really annoying to figure it out. Actually there should be two versions of UCX rpms like in HPCX-MPI where multi-threaded libs are in different directory.
Hi Josif,
Thank you for notifying us about this issue.
Could you please open a support case at support@mellanox.com for further debug?
Thanks,
Samer