V2.4 - cherry-pick (#46126)

* back fl

* delete ssl cert

* .

* make warning

* .

* unittest paral degree

* solve unittest

* heter & multi cloud commm ready

* .

* .

* fix gloo compile warning

* adapt for nn fl-ps

cmake/external/gloo.cmake

@@ -25,8 +25,8 @@ set(GLOO_LIBRARY_DIR
     "${GLOO_INSTALL_DIR}/lib"
     CACHE PATH "gloo library directory." FORCE)
 # As we add extra features for gloo, we use the non-official repo
-set(GLOO_REPOSITORY ${GIT_URL}/sandyhouse/gloo.git)
-set(GLOO_TAG v0.0.2)
+set(GLOO_REPOSITORY ${GIT_URL}/ziyoujiyi/gloo.git)
+set(GLOO_TAG v0.0.3)
 set(GLOO_LIBRARIES
     "${GLOO_INSTALL_DIR}/lib/libgloo.a"
     CACHE FILEPATH "gloo library." FORCE)

python/paddle/distributed/ps/the_one_ps.py

@@ -1090,8 +1090,9 @@ class TheOnePSRuntime(RuntimeBase):
         print("communicator config:", trainer_config.get_communicator_flags())
 
         self._worker.init_worker(worker_desc, self.string_hosts, self.role_id)
-        self.trainer_endpoint = get_trainer_endpoint(self.role_maker)
-        print("fl-ps > trainer_endpoint: {}".format(self.trainer_endpoint))
+        if not self.is_heter_ps_mode:
+            self.trainer_endpoint = get_trainer_endpoint(self.role_maker)
+            print("fl-ps > trainer_endpoint: {}".format(self.trainer_endpoint))
         print("fl-ps > with_coordinator? {}".format(self.with_coordinator))
         print("fl-ps > coordinator addr: {}".format(self.coordinator_hosts))
         if self.with_coordinator: