Replace MSCCLPP_MNNVL_NRANKS_PER_NODE (which overrode TcpBootstrap and
silently changed getNranksPerNode() for every consumer) with a single
algorithm-level helper getIpcDomainNranks(comm) backed by a new
MSCCLPP_IPC_DOMAIN_NRANKS env. The neutral IPC name covers both NVLink/
MNNVL on NV and XGMI on AMD. Bootstrap is unchanged and continues to
report physical-host detection.
Collapse the two getCollectiveDomainNranksPerNode overloads into one
canonical helper and route all six allreduce algos (packet,
allpair_packet, nvls_packet, nvls_zero_copy, rsag, rsag_zero_copy)
through it. Update the standalone tuning example to use the new env
name; drop the undeclared MSCCLPP_ENABLE_MNNVL gate; fix
multi_host_mnnvl detection now that nranks_per_node is no longer
overridden by the bootstrap.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Add an MNNVL rank-domain override so MSCCL++ collectives can treat multi-host NVLink fabrics as a single CUDA IPC/NVLS peer group. Update packet, RSAG, and NVLS allreduce paths to use the collective domain size and teach the torch integration tuning example to select MNNVL-capable allreduce algorithms.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
