revert memcpy use for direct AG (#2146)

Co-authored-by: Islam <nusislam@amd.com>

src/collectives.cc

@@ -131,16 +131,14 @@ ncclResult_t ncclAllGather_impl(const void* sendbuff, void* recvbuff, size_t sen
         dstBuf = recvbuff;
      }
 
-     if (!in_place)
-         CUDACHECK(cudaMemcpyAsync((char*)dstBuf + rank * rankOffset, srcBuf, rankOffset, cudaMemcpyDeviceToDevice, stream));
-     
      NCCLCHECK(ncclGroupStart());
 
      for (int r = 0; r < nRanks; r++) {
-         if (r != rank) {
-             NCCLCHECK(ncclSend(((char*)dstBuf) + rank * rankOffset, sendcount, datatype, r, comm, stream));
-             NCCLCHECK(ncclRecv(((char*)dstBuf) + r * rankOffset, sendcount, datatype, r, comm, stream));
-         }
+         if (r == rank && in_place)
+             continue;
+         
+         NCCLCHECK(ncclSend(((char*)srcBuf), sendcount, datatype, r, comm, stream));
+         NCCLCHECK(ncclRecv(((char*)dstBuf) + r * rankOffset, sendcount, datatype, r, comm, stream));
      }
      NCCLCHECK(ncclGroupEnd());
      return ncclSuccess;