Raise exception for ddp_cpu not supported for TPUs

docs/source/governance.rst

@@ -39,7 +39,7 @@ Board
 
 Alumni
 ------
-- Jeff Yang (`ydcjeff <https://github.com/ydcjeff>`_)   
+- Jeff Yang (`ydcjeff <https://github.com/ydcjeff>`_)
 - Jeff Ling (`jeffling <https://github.com/jeffling>`_)
 - Teddy Koker (`teddykoker <https://github.com/teddykoker>`_)
 - Nate Raw (`nateraw <https://github.com/nateraw>`_)

pytorch_lightning/trainer/connectors/accelerator_connector.py

@@ -742,6 +742,11 @@ def set_distributed_mode(self, distributed_backend: Optional[str] = None):
 
         # special case with DDP on CPUs
         if self.distributed_backend == "ddp_cpu":
+            if _TPU_AVAILABLE:
+                raise MisconfigurationException(
+                    "`accelerator='ddp_cpu'` is not supported on TPU machines. "
+                    "Learn more: https://github.com/PyTorchLightning/pytorch-lightning/issues/7810"
+                )
             self._distrib_type = DistributedType.DDP_SPAWN
             if self.num_gpus > 0:
                 rank_zero_warn(

tests/accelerators/test_tpu_backend.py

@@ -264,3 +264,10 @@ def on_train_end(self, trainer, pl_module):
 
     for param, param_copy in zip(model.parameters(), model_copy.parameters()):
         assert not torch.equal(param.cpu().data, param_copy.data)
+
+
+@RunIf(tpu=True)
+def test_ddp_cpu_not_supported_on_tpus():
+
+    with pytest.raises(MisconfigurationException, match="`accelerator='ddp_cpu'` is not supported on TPU machines"):
+        Trainer(accelerator="ddp_cpu")