Fix MSCCLPP accuracy issue for allreduce7 (#1634)

* ext-src: fix a graph-mode bug in allreduce7

* change MSCCLPP threshold to 16MB

* ext-src: change message size threshold for allreduce7

* ext-src: address review comments

src/init.cc

@@ -114,7 +114,7 @@ bool operator ==(const ncclUniqueId& a, const ncclUniqueId& b) {
   return memcmp(a.internal, b.internal, NCCL_UNIQUE_ID_BYTES) == 0;
 }
 
-RCCL_PARAM(MscclppThreshold, "MSCCLPP_THRESHOLD", (size_t)(1024*1024));
+RCCL_PARAM(MscclppThreshold, "MSCCLPP_THRESHOLD", (size_t)(16*1024*1024));
 #endif
 
 static constexpr int64_t defaultEnableMscclpp = 0;