[TCPStore] Throw value error if passing world_size=0 to TCPStore (p…

…ytorch#137792)

This fixes pytorch#137577.

Pull Request resolved: pytorch#137792
Approved by: https://github.com/fegin, https://github.com/H-Huang
ghstack dependencies: pytorch#137713, pytorch#137721

test/distributed/test_store.py

@@ -516,6 +516,11 @@ def test_store_timeout_on_missing_clients(self):
             use_libuv=self._use_libuv,
         )
 
+    @skip_if_win32()
+    def test_world_size_0_raises(self):
+        with self.assertRaisesRegex(ValueError, "TCPStore world size cannot be 0"):
+            dist.TCPStore("localhost", 0, world_size=0, is_master=False)
+
 
 class LibUvTCPStoreTest(TCPStoreTest):
     _use_libuv = True

torch/csrc/distributed/c10d/init.cpp

@@ -1534,6 +1534,9 @@ Example::
                       bool useLibUV) {
             std::optional<std::size_t> numWorkers = std::nullopt;
             if (worldSize.has_value() && worldSize.value() > -1) {
+              if (worldSize.value() == 0) {
+                throw py::value_error("TCPStore world size cannot be 0");
+              }
               numWorkers = static_cast<std::size_t>(worldSize.value());
             }