2.19.4-1

Split transport connect phase into multiple steps to avoid port
exhaustion when connecting alltoall at large scale. Defaults to 128
peers per round.
Fix memory leaks on CUDA graph capture.
Fix alltoallv crash on self-sendrecv.
Make topology detection more deterministic when PCI speeds are not
available (fix issue #1020).
Properly close shared memory in NVLS resources.
Revert proxy detach after 5 seconds.
Add option to print progress during transport connect.
Add option to set NCCL_DEBUG to INFO on first WARN.

src/graph/xml.cc

@@ -9,6 +9,7 @@
 #include <unistd.h>
 #include <fcntl.h>
 #include <ctype.h>
+#include <float.h>
 #include "core.h"
 #include "nvmlwrap.h"
 #include "xml.h"
@@ -500,11 +501,11 @@ ncclResult_t ncclTopoGetXmlFromSys(struct ncclXmlNode* pciNode, struct ncclXml*
   if (index == -1) {
     if (path) {
       char deviceSpeedStr[MAX_STR_LEN];
-      float deviceSpeed;
+      float deviceSpeed = FLT_MAX;
       NCCLCHECK(ncclTopoGetStrFromSys(path, "max_link_speed", deviceSpeedStr));
       sscanf(deviceSpeedStr, "%f GT/s", &deviceSpeed);
       char portSpeedStr[MAX_STR_LEN];
-      float portSpeed;
+      float portSpeed = FLT_MAX;
       NCCLCHECK(ncclTopoGetStrFromSys(path, "../max_link_speed", portSpeedStr));
       sscanf(portSpeedStr, "%f GT/s", &portSpeed);
       NCCLCHECK(xmlSetAttr(pciNode, "link_speed", portSpeed < deviceSpeed ? portSpeedStr : deviceSpeedStr));