2.10.3-1

Add support for bfloat16.
Add ncclAvg reduction operation.
Improve performance for aggregated operations.
Improve performance for tree.
Improve network error reporting.
Add NCCL_NET parameter to force a specific network.
Add NCCL_IB_QPS_PER_CONNECTION parameter to split IB traffic onto multiple queue pairs.
Fix topology detection error in WSL2.
Fix proxy memory elements affinity (improve alltoall performance).
Fix graph search on cubemesh topologies.
Fix hang in cubemesh during NVB connections.

pkg/redhat/nccl.spec.in

@@ -7,7 +7,7 @@ Group:          Development/Libraries
 License:        BSD
 URL:            http://developer.nvidia.com/nccl
 Source0:        nccl_${nccl:Major}.${nccl:Minor}.${nccl:Patch}${nccl:Suffix}-${pkg:Revision}+cuda${cuda:Major}.${cuda:Minor}_${pkg:Arch}.txz
-Prereq:         /sbin/ldconfig
+Requires(pre,preun): /sbin/ldconfig
 
 %description
 NCCL (pronounced "Nickel") is a stand-alone library of standard collective
@@ -46,6 +46,7 @@ ln -s libnccl.so.${nccl:Major}.${nccl:Minor}.${nccl:Patch} $RPM_BUILD_ROOT/%{_li
 # devel
 install -m 755 -d $RPM_BUILD_ROOT/%{_includedir}
 install -m 644 include/nccl.h $RPM_BUILD_ROOT/%{_includedir}
+install -m 644 include/nccl_net.h $RPM_BUILD_ROOT/%{_includedir}
 ln -s libnccl.so.${nccl:Major} $RPM_BUILD_ROOT/%{_libdir}/libnccl.so
 
 # static
@@ -64,6 +65,7 @@ rm -rf $RPM_BUILD_ROOT
 %doc LICENSE.txt
 %defattr(-,root,root,-)
 %{_includedir}/nccl.h
+%{_includedir}/nccl_net.h
 %{_libdir}/libnccl.so
 
 %files static