test(prim-cinn): split test_resnet and test_bert into three tests (#53723)

* test(prim-cinn): split test_resnet and test_bert into three tests

* test(prim-cinn): fix cmake file to run prim test in CINN-CI

test/prim/model/CMakeLists.txt

@@ -8,12 +8,21 @@ foreach(TEST_OP ${TEST_OPS})
   py_test_modules(${TEST_OP} MODULES ${TEST_OP} ENVS ${GC_ENVS})
 endforeach()
 
-set_tests_properties(test_resnet_prim_cinn PROPERTIES TIMEOUT 850)
-set_tests_properties(test_bert_prim_cinn PROPERTIES TIMEOUT 500)
+set_tests_properties(test_resnet_prim PROPERTIES TIMEOUT 850)
+set_tests_properties(test_bert_prim PROPERTIES TIMEOUT 500)
 set_tests_properties(test_prim_simplenet_cinn PROPERTIES TIMEOUT 120)
 
 if(WITH_CINN)
+  set_tests_properties(test_resnet_cinn PROPERTIES TIMEOUT 850)
+  set_tests_properties(test_resnet_prim_cinn PROPERTIES TIMEOUT 850)
+  set_tests_properties(test_bert_cinn PROPERTIES TIMEOUT 500)
+  set_tests_properties(test_bert_prim_cinn PROPERTIES TIMEOUT 500)
+
+  set_tests_properties(test_resnet_prim PROPERTIES LABELS "RUN_TYPE=CINN")
+  set_tests_properties(test_resnet_cinn PROPERTIES LABELS "RUN_TYPE=CINN")
   set_tests_properties(test_resnet_prim_cinn PROPERTIES LABELS "RUN_TYPE=CINN")
+  set_tests_properties(test_bert_prim PROPERTIES LABELS "RUN_TYPE=CINN")
+  set_tests_properties(test_bert_cinn PROPERTIES LABELS "RUN_TYPE=CINN")
   set_tests_properties(test_bert_prim_cinn PROPERTIES LABELS "RUN_TYPE=CINN")
   set_tests_properties(test_prim_simplenet_cinn PROPERTIES LABELS
                                                            "RUN_TYPE=CINN")

test/prim/model/test_bert_cinn.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import time
+import unittest
+
+import numpy as np
+from bert import Bert, BertPretrainingCriterion, create_pretraining_dataset
+
+import paddle
+from paddle import fluid
+from paddle.dataset.common import DATA_HOME, download
+from paddle.fluid import core
+
+SEED = 2023
+BATCH_SIZE = 2
+
+URL = 'https://paddle-ci.gz.bcebos.com/prim_cinn/bert_training_data.npz'
+MODULE_NAME = 'test_bert_prim_cinn'
+MD5SUM = '71e730ee8d7aa77a215b7e898aa089af'
+SAVE_NAME = 'bert_training_data.npz'
+
+
+DY2ST_CINN_GT = [
+    10.649632453918457,
+    10.333406448364258,
+    10.33541202545166,
+    10.260543823242188,
+    10.219606399536133,
+    10.176884651184082,
+    10.124699592590332,
+    10.072620391845703,
+    10.112163543701172,
+    9.969393730163574,
+]
+
+
+if core.is_compiled_with_cuda():
+    paddle.set_flags({'FLAGS_cudnn_deterministic': True})
+
+
+def train(to_static, enable_prim, enable_cinn):
+    if core.is_compiled_with_cuda():
+        paddle.set_device('gpu')
+    else:
+        paddle.set_device('cpu')
+    fluid.core._set_prim_all_enabled(enable_prim)
+
+    np.random.seed(SEED)
+    paddle.seed(SEED)
+    # paddle.framework.random._manual_program_seed(SEED)
+
+    train_data_loader = create_pretraining_dataset(
+        os.path.join(DATA_HOME, MODULE_NAME, SAVE_NAME),
+        20,
+        {},
+        batch_size=BATCH_SIZE,
+        worker_init=None,
+    )
+
+    # Now only apply dy2st for encoder
+    bert = Bert(to_static, enable_cinn)
+    criterion = BertPretrainingCriterion()
+
+    optimizer = fluid.optimizer.Adam(parameter_list=bert.parameters())
+
+    losses = []
+    for step, batch in enumerate(train_data_loader):
+        start_time = time.time()
+        (
+            input_ids,
+            segment_ids,
+            input_mask,
+            masked_lm_positions,
+            masked_lm_labels,
+            next_sentence_labels,
+            masked_lm_scale,
+        ) = batch
+
+        prediction_scores, seq_relationship_score = bert(
+            input_ids=input_ids,
+            token_type_ids=segment_ids,
+            attention_mask=input_mask,
+            masked_positions=masked_lm_positions,
+        )
+
+        loss = criterion(
+            prediction_scores,
+            seq_relationship_score,
+            masked_lm_labels,
+            next_sentence_labels,
+            masked_lm_scale,
+        )
+
+        loss.backward()
+        optimizer.minimize(loss)
+        bert.clear_gradients()
+        losses.append(loss.numpy().item())
+
+        print(
+            "step: {}, loss: {}, batch_cost: {:.5}".format(
+                step,
+                loss.numpy(),
+                time.time() - start_time,
+            )
+        )
+        if step >= 9:
+            break
+    print(losses)
+    return losses
+
+
+class TestBert(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        download(URL, MODULE_NAME, MD5SUM, SAVE_NAME)
+
+    def tearDown(self):
+        paddle.set_flags({'FLAGS_deny_cinn_ops': ''})
+
+    @unittest.skipIf(
+        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
+        "paddle is not compiled with CINN and CUDA",
+    )
+    def test_cinn(self):
+        paddle.set_flags({'FLAGS_deny_cinn_ops': "dropout"})
+        dy2st_cinn = train(to_static=True, enable_prim=False, enable_cinn=True)
+        np.testing.assert_allclose(dy2st_cinn, DY2ST_CINN_GT, rtol=1e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()

test/prim/model/test_bert_prim.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+import time
+import unittest
+
+import numpy as np
+from bert import Bert, BertPretrainingCriterion, create_pretraining_dataset
+
+import paddle
+from paddle import fluid
+from paddle.dataset.common import DATA_HOME, download
+from paddle.fluid import core
+
+SEED = 2023
+BATCH_SIZE = 2
+
+URL = 'https://paddle-ci.gz.bcebos.com/prim_cinn/bert_training_data.npz'
+MODULE_NAME = 'test_bert_prim_cinn'
+MD5SUM = '71e730ee8d7aa77a215b7e898aa089af'
+SAVE_NAME = 'bert_training_data.npz'
+
+
+DY2ST_PRIM_GT = [
+    11.144556999206543,
+    10.343620300292969,
+    10.330279350280762,
+    10.276118278503418,
+    10.222086906433105,
+    10.194628715515137,
+    10.14902114868164,
+    10.096250534057617,
+    10.104615211486816,
+    9.985644340515137,
+]
+
+if core.is_compiled_with_cuda():
+    paddle.set_flags({'FLAGS_cudnn_deterministic': True})
+
+
+def train(to_static, enable_prim, enable_cinn):
+    if core.is_compiled_with_cuda():
+        paddle.set_device('gpu')
+    else:
+        paddle.set_device('cpu')
+    fluid.core._set_prim_all_enabled(enable_prim)
+
+    np.random.seed(SEED)
+    paddle.seed(SEED)
+    # paddle.framework.random._manual_program_seed(SEED)
+
+    train_data_loader = create_pretraining_dataset(
+        os.path.join(DATA_HOME, MODULE_NAME, SAVE_NAME),
+        20,
+        {},
+        batch_size=BATCH_SIZE,
+        worker_init=None,
+    )
+
+    # Now only apply dy2st for encoder
+    bert = Bert(to_static, enable_cinn)
+    criterion = BertPretrainingCriterion()
+
+    optimizer = fluid.optimizer.Adam(parameter_list=bert.parameters())
+
+    losses = []
+    for step, batch in enumerate(train_data_loader):
+        start_time = time.time()
+        (
+            input_ids,
+            segment_ids,
+            input_mask,
+            masked_lm_positions,
+            masked_lm_labels,
+            next_sentence_labels,
+            masked_lm_scale,
+        ) = batch
+
+        prediction_scores, seq_relationship_score = bert(
+            input_ids=input_ids,
+            token_type_ids=segment_ids,
+            attention_mask=input_mask,
+            masked_positions=masked_lm_positions,
+        )
+
+        loss = criterion(
+            prediction_scores,
+            seq_relationship_score,
+            masked_lm_labels,
+            next_sentence_labels,
+            masked_lm_scale,
+        )
+
+        loss.backward()
+        optimizer.minimize(loss)
+        bert.clear_gradients()
+        losses.append(loss.numpy().item())
+
+        print(
+            "step: {}, loss: {}, batch_cost: {:.5}".format(
+                step,
+                loss.numpy(),
+                time.time() - start_time,
+            )
+        )
+        if step >= 9:
+            break
+    print(losses)
+    return losses
+
+
+class TestBert(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        download(URL, MODULE_NAME, MD5SUM, SAVE_NAME)
+
+    def tearDown(self):
+        paddle.set_flags({'FLAGS_deny_cinn_ops': ''})
+
+    @unittest.skipIf(
+        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
+        "paddle is not compiled with CINN and CUDA",
+    )
+    def test_prim(self):
+        dy2st_prim = train(to_static=True, enable_prim=True, enable_cinn=False)
+        np.testing.assert_allclose(dy2st_prim, DY2ST_PRIM_GT, rtol=1e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()

test/prim/model/test_bert_prim_cinn.py

@@ -33,31 +33,6 @@ MD5SUM = '71e730ee8d7aa77a215b7e898aa089af'
 SAVE_NAME = 'bert_training_data.npz'
 
 
-DY2ST_PRIM_GT = [
-    11.144556999206543,
-    10.343620300292969,
-    10.330279350280762,
-    10.276118278503418,
-    10.222086906433105,
-    10.194628715515137,
-    10.14902114868164,
-    10.096250534057617,
-    10.104615211486816,
-    9.985644340515137,
-]
-DY2ST_CINN_GT = [
-    10.649632453918457,
-    10.333406448364258,
-    10.33541202545166,
-    10.260543823242188,
-    10.219606399536133,
-    10.176884651184082,
-    10.124699592590332,
-    10.072620391845703,
-    10.112163543701172,
-    9.969393730163574,
-]
-
 DY2ST_PRIM_CINN_GT = [
     10.976988792419434,
     10.345282554626465,
@@ -155,23 +130,6 @@ class TestBert(unittest.TestCase):
     def tearDown(self):
         paddle.set_flags({'FLAGS_deny_cinn_ops': ''})
 
-    @unittest.skipIf(
-        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
-        "paddle is not compiled with CINN and CUDA",
-    )
-    def test_prim(self):
-        dy2st_prim = train(to_static=True, enable_prim=True, enable_cinn=False)
-        np.testing.assert_allclose(dy2st_prim, DY2ST_PRIM_GT, rtol=1e-5)
-
-    @unittest.skipIf(
-        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
-        "paddle is not compiled with CINN and CUDA",
-    )
-    def test_cinn(self):
-        paddle.set_flags({'FLAGS_deny_cinn_ops': "dropout"})
-        dy2st_cinn = train(to_static=True, enable_prim=False, enable_cinn=True)
-        np.testing.assert_allclose(dy2st_cinn, DY2ST_CINN_GT, rtol=1e-5)
-
     @unittest.skipIf(
         not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
         "paddle is not compiled with CINN and CUDA",

test/prim/model/test_resnet_cinn.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import unittest
+
+import numpy as np
+
+import paddle
+from paddle import fluid
+from paddle.fluid import core
+from paddle.vision.models import resnet50
+
+SEED = 2020
+base_lr = 0.001
+momentum_rate = 0.9
+l2_decay = 1e-4
+batch_size = 2
+epoch_num = 1
+
+# In V100, 16G, CUDA 11.2, the results are as follows:
+# DY2ST_CINN_GT = [
+#     5.847336769104004,
+#     8.336246490478516,
+#     5.108744144439697,
+#     8.316713333129883,
+#     8.175262451171875,
+#     7.590441703796387,
+#     9.895681381225586,
+#     8.196207046508789,
+#     8.438933372497559,
+#     10.305074691772461,
+
+
+# The results in ci as as follows:
+DY2ST_CINN_GT = [
+    5.828789710998535,
+    8.340764999389648,
+    4.998944282531738,
+    8.474305152893066,
+    8.09157943725586,
+    7.440057754516602,
+    9.907357215881348,
+    8.304681777954102,
+    8.383116722106934,
+    10.120304107666016,
+]
+
+if core.is_compiled_with_cuda():
+    paddle.set_flags({'FLAGS_cudnn_deterministic': True})
+
+
+def reader_decorator(reader):
+    def __reader__():
+        for item in reader():
+            img = np.array(item[0]).astype('float32').reshape(3, 224, 224)
+            label = np.array(item[1]).astype('int64').reshape(1)
+            yield img, label
+
+    return __reader__
+
+
+def optimizer_setting(parameter_list=None):
+    optimizer = fluid.optimizer.Momentum(
+        learning_rate=base_lr,
+        momentum=momentum_rate,
+        regularization=fluid.regularizer.L2Decay(l2_decay),
+        parameter_list=parameter_list,
+    )
+
+    return optimizer
+
+
+def run(model, data_loader, optimizer, mode):
+    if mode == 'train':
+        model.train()
+        end_step = 9
+    elif mode == 'eval':
+        model.eval()
+        end_step = 1
+
+    for epoch in range(epoch_num):
+        total_acc1 = 0.0
+        total_acc5 = 0.0
+        total_sample = 0
+        losses = []
+
+        for batch_id, data in enumerate(data_loader()):
+            start_time = time.time()
+            img, label = data
+
+            pred = model(img)
+            avg_loss = paddle.nn.functional.cross_entropy(
+                input=pred,
+                label=label,
+                soft_label=False,
+                reduction='mean',
+                use_softmax=True,
+            )
+
+            acc_top1 = paddle.static.accuracy(input=pred, label=label, k=1)
+            acc_top5 = paddle.static.accuracy(input=pred, label=label, k=5)
+
+            if mode == 'train':
+                avg_loss.backward()
+                optimizer.minimize(avg_loss)
+                model.clear_gradients()
+
+            total_acc1 += acc_top1
+            total_acc5 += acc_top5
+            total_sample += 1
+            losses.append(avg_loss.numpy().item())
+
+            end_time = time.time()
+            print(
+                "[%s]epoch %d | batch step %d, loss %0.8f, acc1 %0.3f, acc5 %0.3f, time %f"
+                % (
+                    mode,
+                    epoch,
+                    batch_id,
+                    avg_loss,
+                    total_acc1.numpy() / total_sample,
+                    total_acc5.numpy() / total_sample,
+                    end_time - start_time,
+                )
+            )
+            if batch_id >= end_step:
+                # avoid dataloader throw abort signaal
+                data_loader._reset()
+                break
+    print(losses)
+    return losses
+
+
+def train(to_static, enable_prim, enable_cinn):
+    if core.is_compiled_with_cuda():
+        paddle.set_device('gpu')
+    else:
+        paddle.set_device('cpu')
+    np.random.seed(SEED)
+    paddle.seed(SEED)
+    paddle.framework.random._manual_program_seed(SEED)
+    fluid.core._set_prim_all_enabled(enable_prim)
+
+    train_reader = paddle.batch(
+        reader_decorator(paddle.dataset.flowers.train(use_xmap=False)),
+        batch_size=batch_size,
+        drop_last=True,
+    )
+    data_loader = fluid.io.DataLoader.from_generator(capacity=5, iterable=True)
+    data_loader.set_sample_list_generator(train_reader)
+
+    resnet = resnet50(False)
+    if to_static:
+        build_strategy = paddle.static.BuildStrategy()
+        if enable_cinn:
+            build_strategy.build_cinn_pass = True
+        resnet = paddle.jit.to_static(resnet, build_strategy=build_strategy)
+    optimizer = optimizer_setting(parameter_list=resnet.parameters())
+
+    train_losses = run(resnet, data_loader, optimizer, 'train')
+    if to_static and enable_prim and enable_cinn:
+        eval_losses = run(resnet, data_loader, optimizer, 'eval')
+    return train_losses
+
+
+class TestResnet(unittest.TestCase):
+    @unittest.skipIf(
+        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
+        "paddle is not compiled with CINN and CUDA",
+    )
+    def test_cinn(self):
+        dy2st_cinn = train(to_static=True, enable_prim=False, enable_cinn=True)
+        np.testing.assert_allclose(dy2st_cinn, DY2ST_CINN_GT, rtol=1e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()

test/prim/model/test_resnet_prim.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import time
+import unittest
+
+import numpy as np
+
+import paddle
+from paddle import fluid
+from paddle.fluid import core
+from paddle.vision.models import resnet50
+
+SEED = 2020
+base_lr = 0.001
+momentum_rate = 0.9
+l2_decay = 1e-4
+batch_size = 2
+epoch_num = 1
+
+# In V100, 16G, CUDA 11.2, the results are as follows:
+# DY2ST_PRIM_GT = [
+#     5.8473358154296875,
+#     8.354944229125977,
+#     5.098367691040039,
+#     8.533346176147461,
+#     8.179085731506348,
+#     7.285282135009766,
+#     9.824585914611816,
+#     8.56928825378418,
+#     8.539499282836914,
+#     10.256929397583008,
+# ]
+
+# The results in ci as as follows:
+DY2ST_PRIM_GT = [
+    5.82879114151001,
+    8.33370590209961,
+    5.091761589050293,
+    8.776082992553711,
+    8.274380683898926,
+    7.546653747558594,
+    9.607137680053711,
+    8.27371597290039,
+    8.429732322692871,
+    10.362630844116211,
+]
+
+if core.is_compiled_with_cuda():
+    paddle.set_flags({'FLAGS_cudnn_deterministic': True})
+
+
+def reader_decorator(reader):
+    def __reader__():
+        for item in reader():
+            img = np.array(item[0]).astype('float32').reshape(3, 224, 224)
+            label = np.array(item[1]).astype('int64').reshape(1)
+            yield img, label
+
+    return __reader__
+
+
+def optimizer_setting(parameter_list=None):
+    optimizer = fluid.optimizer.Momentum(
+        learning_rate=base_lr,
+        momentum=momentum_rate,
+        regularization=fluid.regularizer.L2Decay(l2_decay),
+        parameter_list=parameter_list,
+    )
+
+    return optimizer
+
+
+def run(model, data_loader, optimizer, mode):
+    if mode == 'train':
+        model.train()
+        end_step = 9
+    elif mode == 'eval':
+        model.eval()
+        end_step = 1
+
+    for epoch in range(epoch_num):
+        total_acc1 = 0.0
+        total_acc5 = 0.0
+        total_sample = 0
+        losses = []
+
+        for batch_id, data in enumerate(data_loader()):
+            start_time = time.time()
+            img, label = data
+
+            pred = model(img)
+            avg_loss = paddle.nn.functional.cross_entropy(
+                input=pred,
+                label=label,
+                soft_label=False,
+                reduction='mean',
+                use_softmax=True,
+            )
+
+            acc_top1 = paddle.static.accuracy(input=pred, label=label, k=1)
+            acc_top5 = paddle.static.accuracy(input=pred, label=label, k=5)
+
+            if mode == 'train':
+                avg_loss.backward()
+                optimizer.minimize(avg_loss)
+                model.clear_gradients()
+
+            total_acc1 += acc_top1
+            total_acc5 += acc_top5
+            total_sample += 1
+            losses.append(avg_loss.numpy().item())
+
+            end_time = time.time()
+            print(
+                "[%s]epoch %d | batch step %d, loss %0.8f, acc1 %0.3f, acc5 %0.3f, time %f"
+                % (
+                    mode,
+                    epoch,
+                    batch_id,
+                    avg_loss,
+                    total_acc1.numpy() / total_sample,
+                    total_acc5.numpy() / total_sample,
+                    end_time - start_time,
+                )
+            )
+            if batch_id >= end_step:
+                # avoid dataloader throw abort signaal
+                data_loader._reset()
+                break
+    print(losses)
+    return losses
+
+
+def train(to_static, enable_prim, enable_cinn):
+    if core.is_compiled_with_cuda():
+        paddle.set_device('gpu')
+    else:
+        paddle.set_device('cpu')
+    np.random.seed(SEED)
+    paddle.seed(SEED)
+    paddle.framework.random._manual_program_seed(SEED)
+    fluid.core._set_prim_all_enabled(enable_prim)
+
+    train_reader = paddle.batch(
+        reader_decorator(paddle.dataset.flowers.train(use_xmap=False)),
+        batch_size=batch_size,
+        drop_last=True,
+    )
+    data_loader = fluid.io.DataLoader.from_generator(capacity=5, iterable=True)
+    data_loader.set_sample_list_generator(train_reader)
+
+    resnet = resnet50(False)
+    if to_static:
+        build_strategy = paddle.static.BuildStrategy()
+        if enable_cinn:
+            build_strategy.build_cinn_pass = True
+        resnet = paddle.jit.to_static(resnet, build_strategy=build_strategy)
+    optimizer = optimizer_setting(parameter_list=resnet.parameters())
+
+    train_losses = run(resnet, data_loader, optimizer, 'train')
+    if to_static and enable_prim and enable_cinn:
+        eval_losses = run(resnet, data_loader, optimizer, 'eval')
+    return train_losses
+
+
+class TestResnet(unittest.TestCase):
+    @unittest.skipIf(
+        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
+        "paddle is not compiled with CINN and CUDA",
+    )
+    def test_prim(self):
+        dy2st_prim = train(to_static=True, enable_prim=True, enable_cinn=False)
+        np.testing.assert_allclose(dy2st_prim, DY2ST_PRIM_GT, rtol=1e-5)
+
+
+if __name__ == '__main__':
+    unittest.main()

test/prim/model/test_resnet_prim_cinn.py

@@ -30,30 +30,6 @@ batch_size = 2
 epoch_num = 1
 
 # In V100, 16G, CUDA 11.2, the results are as follows:
-# DY2ST_PRIM_GT = [
-#     5.8473358154296875,
-#     8.354944229125977,
-#     5.098367691040039,
-#     8.533346176147461,
-#     8.179085731506348,
-#     7.285282135009766,
-#     9.824585914611816,
-#     8.56928825378418,
-#     8.539499282836914,
-#     10.256929397583008,
-# ]
-# DY2ST_CINN_GT = [
-#     5.847336769104004,
-#     8.336246490478516,
-#     5.108744144439697,
-#     8.316713333129883,
-#     8.175262451171875,
-#     7.590441703796387,
-#     9.895681381225586,
-#     8.196207046508789,
-#     8.438933372497559,
-#     10.305074691772461,
-# ]
 # DY2ST_PRIM_CINN_GT = [
 #     5.8473358154296875,
 #     8.322463989257812,
@@ -68,31 +44,6 @@ epoch_num = 1
 # ]
 
 # The results in ci as as follows:
-DY2ST_PRIM_GT = [
-    5.82879114151001,
-    8.33370590209961,
-    5.091761589050293,
-    8.776082992553711,
-    8.274380683898926,
-    7.546653747558594,
-    9.607137680053711,
-    8.27371597290039,
-    8.429732322692871,
-    10.362630844116211,
-]
-DY2ST_CINN_GT = [
-    5.828789710998535,
-    8.340764999389648,
-    4.998944282531738,
-    8.474305152893066,
-    8.09157943725586,
-    7.440057754516602,
-    9.907357215881348,
-    8.304681777954102,
-    8.383116722106934,
-    10.120304107666016,
-]
-
 DY2ST_PRIM_CINN_GT = [
     5.828786849975586,
     8.332868576049805,
@@ -225,22 +176,6 @@ def train(to_static, enable_prim, enable_cinn):
 
 
 class TestResnet(unittest.TestCase):
-    @unittest.skipIf(
-        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
-        "paddle is not compiled with CINN and CUDA",
-    )
-    def test_prim(self):
-        dy2st_prim = train(to_static=True, enable_prim=True, enable_cinn=False)
-        np.testing.assert_allclose(dy2st_prim, DY2ST_PRIM_GT, rtol=1e-5)
-
-    @unittest.skipIf(
-        not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
-        "paddle is not compiled with CINN and CUDA",
-    )
-    def test_cinn(self):
-        dy2st_cinn = train(to_static=True, enable_prim=False, enable_cinn=True)
-        np.testing.assert_allclose(dy2st_cinn, DY2ST_CINN_GT, rtol=1e-5)
-
     @unittest.skipIf(
         not (paddle.is_compiled_with_cinn() and paddle.is_compiled_with_cuda()),
         "paddle is not compiled with CINN and CUDA",