skip failing tests on ROCm

tests/L0/run_optimizers/test_fused_optimizer.py

@@ -87,6 +87,7 @@ class TestFusedAdam(TestFusedOptimizer):
         self.ref_optim = torch.optim.Adam
         self.fused_optim = apex.optimizers.FusedAdam
 
+    @skipIfRocm
     def test_float(self):
         self.gen_single_type_test(param_type=torch.float)
 
@@ -102,6 +103,7 @@ class TestFusedAdam(TestFusedOptimizer):
         self.max_abs_diff = 1e-2
         self.gen_single_type_test(param_type=torch.bfloat16, apex_only=True)
 
+    @skipIfRocm
     @unittest.skipIf(torch.cuda.device_count()<2, "more than 1 GPU required")
     def test_multi_device(self):
         devices = ("cuda:0", "cuda:1")
@@ -194,6 +196,7 @@ class TestFusedAdagrad(TestFusedOptimizer):
         self.ref_optim = torch.optim.Adagrad
         self.fused_optim = apex.optimizers.FusedAdagrad
 
+    @skipIfRocm
     def test_float(self):
         self.gen_single_type_test(param_type=torch.float)
 
@@ -201,6 +204,7 @@ class TestFusedAdagrad(TestFusedOptimizer):
     def test_half(self):
         self.gen_single_type_test(param_type=torch.float16)
 
+    @skipIfRocm
     @unittest.skipIf(torch.cuda.device_count()<2, "more than 1 GPU required")
     def test_multi_device(self):
         devices = ("cuda:0", "cuda:1")
@@ -209,6 +213,7 @@ class TestFusedAdagrad(TestFusedOptimizer):
                 self.gen_single_type_test(param_type=torch.float, device=tensor_dev)
 
 
+    @skipIfRocm
     def test_multi_params(self):
         sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]
         adagrad_option = {"lr": 5e-4, "eps": 1e-08, "weight_decay": 0}

tests/L0/run_optimizers/test_lamb.py

@@ -228,6 +228,7 @@ class TestFusedLAMB(unittest.TestCase):
             with torch.cuda.device(current_dev):
                 self.gen_single_type_test(param_type=torch.float, device=tensor_dev)
 
+    @skipIfRocm
     def test_multi_params(self):
         sizes = [[4096, 1024], [4096], [4096, 2048], [32320, 1024], [1]]
         weight_decay = [0, 0.01]