Disable broken tests (#1055)

tests/experimental/nn/test_ssd_offload.py

@@ -79,6 +79,7 @@ def test_ssd_handle_dispatch_bwd():
         assert torch.equal(ssd_handle.grad, orig_copy.grad)
 
 
+@pytest.mark.skip("broken at head")
 def test_ssd_handle_dispatch_bwd_hook():
     _init()
 
@@ -277,6 +278,7 @@ def test_ssd_flat_parameter_view_modify():
         assert ssd_flat_param.storage_state == so.StorageState.ON_CPU_DIRTY
 
 
+@pytest.mark.skip("broken at head")
 def test_ssd_flat_parameter_view_bwd():
     _init()
 
@@ -344,6 +346,7 @@ def test_ssd_flat_parameter_view_bwd():
         assert "one" in hooks_called
 
 
+@pytest.mark.skip("broken at head")
 def test_ssd_flat_parameter_view_bwd_parameterization():
     _init()
 

tests/experimental/nn/test_sync_batchnorm.py

@@ -97,7 +97,12 @@ def parity3d_bn():
     check_parity(torch_bn, fs_bn, x)
 
 
-@pg_test()
+@pytest.mark.skip("broken at head")
+def test_parity3d_checkpoint_syncbn():
+    assert 1 == 2
+
+
+# @pg_test()
 def parity3d_checkpoint_syncbn():
     rank = dist.get_rank()
     torch.cuda.set_device(rank)
@@ -110,7 +115,12 @@ def parity3d_checkpoint_syncbn():
     check_parity_ddp(torch_bn, fs_bn, x)
 
 
-@pg_test()
+@pytest.mark.skip("broken at head")
+def test_parity3d_checkpoint_syncbn_twice():
+    assert 1 == 2
+
+
+# @pg_test()
 def parity3d_checkpoint_syncbn_twice():
     rank = dist.get_rank()
     torch.cuda.set_device(rank)

tests/nn/data_parallel/test_fsdp_memory.py

@@ -159,6 +159,7 @@ def _distributed_worker(
 
 
 @skip_if_single_gpu
+@pytest.mark.timeout(120)
 @pytest.mark.parametrize("ckpt", ["no_ckpt", "ckpt"])
 @pytest.mark.parametrize("fsdp", ["ddp", "fsdp", "fsdp_amp_default", "fsdp_amp_compute_dtype32"])
 def test_fsdp_memory(fsdp, ckpt):

tests/nn/data_parallel/test_fsdp_regnet.py

@@ -349,6 +349,7 @@ def _distributed_worker(
 
 # We use strings for precision and flatten params instead of bool to
 # make the pytest output more readable.
+@pytest.mark.skip("broken at head")
 @skip_if_single_gpu
 @pytest.mark.parametrize("precision", ["full", "mixed"])
 @pytest.mark.parametrize("flatten", ["flatten", "no_flatten"])

tests/optim/test_oss.py

@@ -958,6 +958,7 @@ def run_ddp_parity(rank, world_size, backend, temp_file_name, change_train_graph
     dist.destroy_process_group()
 
 
+@pytest.mark.skip("broken at head")
 @skip_if_no_cuda
 @skip_if_single_gpu
 @pytest.mark.parametrize("change_train_graph", [True, False])