2.16.5-1

Add support for 400Gbit NDR network adapters (CX7)
Handle EINTR in socket poll() function
Add NCCL_PROGRESS_APPENDOP_FREQ to control op append overhead
Resource cleanup fixes
Fix double free in case of init failure
Fix crash in ncclCommAbort
Revert AMD speed commit

src/misc/param.cc

@@ -72,7 +72,7 @@ void ncclLoadParam(char const* env, int64_t deftVal, int64_t uninitialized, int6
         value = deftVal;
         INFO(NCCL_ALL,"Invalid value %s for %s, using default %lld.", str, env, (long long)deftVal);
       } else {
-        INFO(NCCL_ALL,"%s set by environment to %lld.", env, (long long)value);
+        INFO(NCCL_ENV,"%s set by environment to %lld.", env, (long long)value);
       }
     }
     __atomic_store_n(cache, value, __ATOMIC_RELAXED);

src/misc/shmutils.cc

@@ -54,7 +54,7 @@ ncclResult_t ncclShmOpen(char* shmPath, size_t shmSize, void** shmPtr, void** de
   const size_t realShmSize = shmSize + refSize;
 
   *handle = *shmPtr = NULL; /* assume shmPtr and handle always set correctly by users. */
-  EQCHECKGOTO(tmphandle = (struct shmHandleInternal*)malloc(sizeof(struct shmHandleInternal)), NULL, ret, fail);
+  EQCHECKGOTO(tmphandle = (struct shmHandleInternal*)calloc(1, sizeof(struct shmHandleInternal)), NULL, ret, fail);
   if (create) {
     /* refcount > 0 means the caller tries to allocate a shared memory. This shared memory segment will have
      * refcount references; when the peer attaches, it should pass -1 to reduce one reference count. When it

src/misc/socket.cc

@@ -490,11 +490,18 @@ static ncclResult_t socketPollConnect(struct ncclSocket* sock) {
   memset(&pfd, 0, sizeof(struct pollfd));
   pfd.fd = sock->fd;
   pfd.events = POLLOUT;
-  SYSCHECK(ret = poll(&pfd, 1, timeout), "poll");
-  if (ret == 0) return ncclSuccess;
+  ret = poll(&pfd, 1, timeout);
+
+  if (ret == 0 || (ret < 0 && errno == EINTR)) {
+    return ncclSuccess;
+  } else if (ret < 0) {
+    WARN("socketPollConnect poll() failed with error %s", strerror(errno));
+    return ncclRemoteError;
+  } else {
+    EQCHECK(ret == 1 && (pfd.revents & POLLOUT), 0);
+  }
 
   /* check socket status */
-  EQCHECK(ret == 1 && (pfd.revents & POLLOUT), 0);
   SYSCHECK(getsockopt(sock->fd, SOL_SOCKET, SO_ERROR, (void*)&ret, &rlen), "getsockopt");
 
   if (ret == 0) {