[BugFix] fix bugs in DCU unit tests (#54874)

* block bf16 tests on ROCM

* block more bf16 tests on ROCM

* some unittest cases doesn't have kernels on ROCm

* some unittest cases doesn't have kernels on ROCm

* fix code style

test/legacy_test/test_assign_op.py

@@ -72,7 +72,8 @@ class TestAssignFP16Op(eager_op_test.OpTest):
 
 
 @unittest.skipIf(
-    not paddle.is_compiled_with_cuda(), "BFP16 test runs only on GPU"
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
 )
 class TestAssignBFP16Op(eager_op_test.OpTest):
     def setUp(self):

test/legacy_test/test_cast_op.py

@@ -95,6 +95,10 @@ class TestCastOpFp32ToFp16(OpTest):
         self.check_grad(['X'], ['Out'], check_prim=True, only_check_prim=True)
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
+)
 class TestCastOpBf16ToFp32(OpTest):
     def setUp(self):
         ipt = np.array(np.random.randint(10, size=[10, 10])).astype('uint16')
@@ -120,6 +124,10 @@ class TestCastOpBf16ToFp32(OpTest):
         self.check_grad(['X'], ['Out'], check_prim=True, only_check_prim=True)
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
+)
 class TestCastOpFp32ToBf16(OpTest):
     def setUp(self):
         ipt = np.random.random(size=[10, 10]).astype('float32')

test/legacy_test/test_elementwise_mul_op.py

@@ -163,6 +163,10 @@ class TestElementwiseMulOp_ZeroDim3(ElementwiseMulOp):
         self.out = np.multiply(self.x, self.y)
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
+)
 class TestBF16ElementwiseMulOp(OpTest):
     def setUp(self):
         self.op_type = "elementwise_mul"

test/legacy_test/test_elementwise_pow_op.py

@@ -268,6 +268,10 @@ class TestElementwisePowOpFP16(OpTest):
         )
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
+)
 class TestElementwisePowBF16Op(OpTest):
     def setUp(self):
         self.op_type = "elementwise_pow"

test/legacy_test/test_fill_any_like_op.py

@@ -64,7 +64,8 @@ class TestFillAnyLikeOpFloat32(TestFillAnyLikeOp):
 
 
 @unittest.skipIf(
-    not core.is_compiled_with_cuda(), "core is not compiled with CUDA"
+    not core.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "core is not compiled with CUDA",
 )
 class TestFillAnyLikeOpBfloat16(OpTest):
     def setUp(self):

test/legacy_test/test_layer_norm_op.py

@@ -126,6 +126,10 @@ def layer_norm_wrapper(
     )
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support fp64 LayerNormOpByOp currently",
+)
 class TestLayerNormOpByOpTest(OpTest):
     def setUp(self):
         self.python_api = layer_norm_wrapper
@@ -164,7 +168,7 @@ class TestLayerNormOpByOpTest(OpTest):
         self.cinn_rtol = 1e-5
 
         self.max_relative_error = 1e-5
-
+        # ROCm does not have float64 LayerNorm kernel
         self.dtype = "float64"
         self.x_shape = [2, 6, 6, 3]
         self.epsilon = 0.00001
@@ -218,6 +222,7 @@ class TestLayerNormOpByOpTest(OpTest):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )
@@ -306,6 +311,10 @@ class TestLayerNormBF16OpByOpTest(OpTest):
         }
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support fp64 LayerNormOpByOp currently",
+)
 class TestLayerNormOpByOpTestFP64_case2(TestLayerNormOpByOpTest):
     def initConfig(self):
         self.rev_comp_atol = 1e-6
@@ -328,6 +337,10 @@ class TestLayerNormOpByOpTestFP64_case2(TestLayerNormOpByOpTest):
         self.has_bias = False
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support bf16 LayerNormOpByOp currently",
+)
 class TestLayerNormBF16OpByOpTest_case2(TestLayerNormBF16OpByOpTest):
     def initConfig(self):
         self.ori_atol = 1e-2
@@ -343,6 +356,10 @@ class TestLayerNormBF16OpByOpTest_case2(TestLayerNormBF16OpByOpTest):
         self.has_bias = False
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support fp64 LayerNormOpByOp currently",
+)
 class TestLayerNormOpByOpTestFP64_case3(TestLayerNormOpByOpTest):
     def initConfig(self):
         self.rev_comp_atol = 1e-7
@@ -365,6 +382,10 @@ class TestLayerNormOpByOpTestFP64_case3(TestLayerNormOpByOpTest):
         self.has_bias = False
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support bf16 LayerNormOpByOp currently",
+)
 class TestLayerNormBF16OpByOpTest_case3(TestLayerNormBF16OpByOpTest):
     def initConfig(self):
         self.ori_atol = 1e-2
@@ -380,6 +401,10 @@ class TestLayerNormBF16OpByOpTest_case3(TestLayerNormBF16OpByOpTest):
         self.has_bias = False
 
 
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(),
+    "ROCm doesn't support fp64 LayerNormOpByOp currently",
+)
 class TestLayerNormOpByOpTestFP64_case4(TestLayerNormOpByOpTest):
     def initConfig(self):
         self.rev_comp_atol = 1e-6
@@ -801,6 +826,10 @@ class TestFP16ScaleBiasLayerNorm(unittest.TestCase):
         assert_equal(b_g_np_1, b_g_np_2)
 
 
+@unittest.skipIf(
+    not core.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BF16 is only supported on CUDA.",
+)
 class TestBF16ScaleBiasLayerNorm(unittest.TestCase):
     def check_main(self, x_np, weight_np, bias_np, dtype):
         paddle.disable_static()
@@ -934,7 +963,7 @@ class TestFastMathLayerNormOp(unittest.TestCase):
         )
 
     def test_main(self):
-        if not paddle.is_compiled_with_cuda():
+        if not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm():
             return
         self.check_with_dtype(dtype="float32")
         self.check_with_dtype(dtype="bfloat16")

test/legacy_test/test_matmul_v2_op.py

@@ -405,6 +405,7 @@ create_test_fp16_class(TestMatMulOpBroadcast2)
 def create_test_bf16_class(parent, atol=0.01):
     @unittest.skipIf(
         not core.is_compiled_with_cuda()
+        or paddle.is_compiled_with_rocm()
         or not core.is_bfloat16_supported(core.CUDAPlace(0)),
         "core is not compiled with CUDA and not support the bfloat16",
     )

test/legacy_test/test_reduce_op.py

@@ -198,7 +198,8 @@ create_test_fp16_class(TestSumOp3Dim)
 
 def create_test_bf16_class(parent):
     @unittest.skipIf(
-        not core.is_compiled_with_cuda(), "core is not compiled with CUDA"
+        not core.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+        "core is not compiled with CUDA",
     )
     class TestSumOpBf16(parent):
         def setUp(self):
@@ -349,6 +350,7 @@ class TestMaxFP16Op(TestMaxFP32Op):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )
@@ -449,6 +451,9 @@ class TestMin8DOp(OpTest):
     reason="reduce_min is discontinuous non-derivable function,"
     " its gradient check is not supported by unittest framework."
 )
+@unittest.skipIf(
+    paddle.is_compiled_with_rocm(), "ROCm doesn't have FP16 reduce_min kernel"
+)
 class TestMinFP16Op(OpTest):
     """Remove Min with subgradient from gradient check to confirm the success of CI."""
 
@@ -479,6 +484,7 @@ class TestMinFP16Op(OpTest):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )
@@ -541,6 +547,7 @@ class TestProdFP16OP(TestProdOp):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )
@@ -648,6 +655,7 @@ class TestProd6DFP16OP(TestProd6DOp):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )
@@ -721,6 +729,7 @@ class TestProd8DFP16OP(TestProd8DOp):
 
 @unittest.skipIf(
     not core.is_compiled_with_cuda()
+    or paddle.is_compiled_with_rocm()
     or not core.is_bfloat16_supported(core.CUDAPlace(0)),
     "core is not compiled with CUDA or not support the bfloat16",
 )

test/legacy_test/test_reshape_op.py

@@ -86,6 +86,10 @@ class TestReshapeOp_ZeroDim3(OpTest):
         self.infered_shape = ()
 
 
+@unittest.skipIf(
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
+)
 class TestReshapeBF16Op(OpTest):
     def setUp(self):
         self.init_data()

test/legacy_test/test_scale_op.py

@@ -155,7 +155,8 @@ class TestScaleFp16Op(TestScaleOp):
 
 
 @unittest.skipIf(
-    not core.is_compiled_with_rocm(), "core is not compiled with CUDA"
+    not paddle.is_compiled_with_cuda() or paddle.is_compiled_with_rocm(),
+    "BFP16 test runs only on CUDA",
 )
 class TestScaleBF16Op(OpTest):
     def setUp(self):

test/legacy_test/test_static_model_parallel_fused_multi_transformer.py

@@ -34,7 +34,10 @@ class TestStaticModelParallel(TestDistBase):
     def test_dist_static_model_parallel_fused_multi_transformer(self):
         from paddle import fluid
 
-        if fluid.core.is_compiled_with_cuda():
+        if (
+            fluid.core.is_compiled_with_cuda()
+            and not paddle.is_compiled_with_rocm()
+        ):
             self.check_with_place(
                 "static_model_parallel_fused_multi_transformer.py",
                 delta=1e-5,