don't set CUDA_MODULE_LOADING (#158712)

If needed, it'll be set in `_C._cuda_init()`. setenv is not threadsafe, so this can cause segfaults due to getenv/setenv races.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/158712
Approved by: https://github.com/eqy

test/test_cuda.py

@@ -6484,11 +6484,6 @@ class TestCudaAutocast(TestAutocast):
             with torch.cuda.amp.autocast():
                 _ = torch.ones(10)
 
-    def test_cuda_module_loading_env(self):
-        torch.cuda.init()
-        val = os.environ.get("CUDA_MODULE_LOADING", "")
-        self.assertEqual(val, "LAZY")
-
 
 @unittest.skipIf(
     os.environ.get("USE_LEGACY_DRIVER", None) == "1", "Doesn't work with older driver"

torch/cuda/__init__.py

@@ -379,8 +379,6 @@ def _lazy_init():
             )
         # This function throws if there's a driver initialization error, no GPUs
         # are found or any other error occurs
-        if "CUDA_MODULE_LOADING" not in os.environ:
-            os.environ["CUDA_MODULE_LOADING"] = "LAZY"
         torch._C._cuda_init()
         # Some of the queued calls may reentrantly call _lazy_init();
         # we need to just return without initializing in that case.