Rebasing nccl 2.28 with latest changes.

Summary:
These diff stack, implements the required updates to rebase NCCL 2.28 for NCCLX with CTRAN integration. It incorporates all changes introduced in version 2.27, applying them on top of NCCL’s latest stable release (2.28). The primary objective of this diff is to enable CTRAN support for NCCLX under NCCL 2.28, ensuring compatibility and leveraging the latest enhancements from the upstream release. In addition, it includes necessary checks inside FastInitTest so that it can ensure the ctran enabling as expected.

Specifically in this diff, all rebasing tasks are implemented.

Reviewed By: zhiyongww

Differential Revision: D85349488

fbshipit-source-id: 370f26f555a224db86c35e03dcf4f4133460c52b

Author: mlunar-meta

comms/ncclx/v2_27/meta/commstate/FactoryCommStateX.cc

@@ -27,9 +27,9 @@ void initRankTopologyFrom(CommStateX* _CommStateX, void* _comm) {
     // each node is already ordered by local rank
     // NOTE: two GPUs on the same node may be with different nodeId because
     // they don't have direct NVL access. To keep same nodeId in statex, we
-    // use hostname+nodeId to make it unique
+    // use hostHash+nodeId to make it unique
     std::string host(
-        std::string(comm->peerInfo[r].hostname) +
+        std::to_string(comm->peerInfo[r].hostHash) + "_" +
         std::to_string(rankState.nodeId));
     _CommStateX->hostToRanks_[host].emplace_back(r);
 

comms/ncclx/v2_27/meta/tests/SimpleCtranInitTest.cc

@@ -14,6 +14,8 @@ class SimpleCtranInitTest : public ::testing::Test {
     setenv("NCCL_CTRAN_ENABLE", "1", 1);
     setenv("NCCL_COLLTRACE", "trace", 1);
     setenv("NCCL_USE_MEM_CACHE", "1", 1);
+    setenv("NCCL_LAZY_SETUP_CHANNELS", "1", 1);
+    setenv("NCCL_RUNTIME_CONNECT", "1", 1);
 
     ncclCvarInit();
     NCCLCHECK_TEST(ncclCudaLibraryInit());