2.18.1-1

Add support for IB SHARP to NVLS (NVLink SHARP algorithm).
Add NVLS+Tree algorithm.
Add support for memory management using cuMem* functions.
Use all NICs for Send/Receive operations on systems with more than
one NIC per GPU (#804).
Add ncclCommSplit primitive, with resource sharing option in config.
Fix alltoallv hang (#788)
Increase number of channels on H100 when we're not limited by NVLink.
Improve error reporting in case of IB failure, printing local and
remote ID (#779).
Add build option to allow compilation against RDMA includes instead
of dynamically loading IB verbs symbols (#802).
Fix context creation for progress thread (#803).
NET/IB: add option to use multiple QPs in round-robin mode.
Fix tree performance issue when NVB is disabled on HCM topologies.

src/misc/shmutils.cc

@@ -14,6 +14,7 @@
 #include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
+#include <utils.h>
 
 struct shmHandleInternal {
   int fd;
@@ -86,17 +87,13 @@ ncclResult_t ncclShmOpen(char* shmPath, size_t shmSize, void** shmPtr, void** de
   if (create) {
     *(int*)(hptr + shmSize) = refcount;
   } else {
-    int remref = __atomic_sub_fetch((int*)(hptr + shmSize), 1, __ATOMIC_RELAXED);
+    int remref = ncclAtomicRefCountDecrement((int*)(hptr + shmSize));
     if (remref == 0) {
       /* the last peer has completed attachment, it should unlink the shm mem file. */
       if (unlink(shmPath) != 0) {
         WARN("unlink shared memory %s failed, error: %s", shmPath, strerror(errno));
       }
     }
-
-    if (refcount != -1) {
-      WARN("attaching memory should only reduce refcount by 1 but %d is passed", refcount);
-    }
   }
 
   if (devShmPtr) {
@@ -133,8 +130,8 @@ ncclResult_t ncclShmClose(ncclShmHandle_t handle) {
           WARN("unlink shared memory %s failed, error: %s", tmphandle->shmPath, strerror(errno));
           ret = ncclSystemError;
         }
-        free(tmphandle->shmPath);
       }
+      free(tmphandle->shmPath);
     }
 
     if (tmphandle->shmPtr) {