!17115 Delete key 0 if all processes in a group are killed

Merge pull request !17115 from chuboning/cherry-pick-1735375236
Ascend · Jan 15, 2025 · 4c396ab · 4c396ab
1 parent b0b2938
commit 4c396ab
Showing 1 changed file with 2 additions and 0 deletions.
diff --git a/torch_npu/csrc/distributed/ProcessGroupHCCL.cpp b/torch_npu/csrc/distributed/ProcessGroupHCCL.cpp
@@ -856,6 +856,8 @@ void ProcessGroupHCCL::abort(c10::optional<std::string> abortReason)
 void ProcessGroupHCCL::deleteTCPStoreKey()
 {
     try {
+        // all processes in a group may be killed, so delete key 0 as a last resort
+        store_->deleteKey("0");
         for (const auto &key : TCPStoreKeyList_) {
             store_->deleteKey(key);
         }