Use inductor TestCase for distributed tests (pytorch#129494)

Summary: At least some of the tests deriving from MultiProcessTestCase exercise inductor. Using the inductor TestCase class makes sure we always get a clean cache dir.

Pull Request resolved: pytorch#129494
Approved by: https://github.com/eellison

Author: masnesral

test/distributed/test_c10d_nccl.py

@@ -2111,12 +2111,12 @@ def setUp(self):
         self._spawn_processes()
 
     def tearDown(self):
-        super().tearDown()
         del os.environ["TORCH_NCCL_ENABLE_TIMING"]
         try:
             os.remove(self.file_name)
         except OSError:
             pass
+        super().tearDown()
 
     def _get_store(self):
         return dist.FileStore(self.file_name, self.world_size)

torch/testing/_internal/common_distributed.py

@@ -29,6 +29,7 @@
 import torch.cuda.nccl
 import torch.distributed as c10d
 import torch.nn as nn
+from torch._inductor.test_case import TestCase, run_tests
 from torch.testing._internal.common_utils import (
     FILE_SCHEMA,
     find_free_port,
@@ -38,8 +39,6 @@
     skip_but_pass_in_sandcastle_if,
     TEST_WITH_ROCM,
     TEST_WITH_TSAN,
-    TestCase,
-    run_tests,
 )
 from torch.testing._internal.distributed.multi_threaded_pg import (
     _install_threaded_pg,