[IPU] add mixed-precission support for ipu (#41733) (#41906)

add mixed-precission support for ipu

cherry-pick from #41733

paddle/fluid/framework/ir/ipu/optimizer_extract_pass.cc

@@ -32,7 +32,8 @@ std::set<std::string> ignored_ops = {
     "elementwise_mul",
     "scale",            // adamax
     "assign",           // adamw
-    "squared_l2_norm"  // gradient_clip_norm
+    "squared_l2_norm",  // gradient_clip_norm
+    "cast",             // mix-precision support
 };
 
 const bool startswith(const std::string& str, const std::string& pre) {

python/paddle/fluid/contrib/mixed_precision/fp16_utils.py

@@ -191,7 +191,8 @@ def _insert_cast_op(block, op, idx, src_dtype, dest_dtype):
                         attrs={
                             "in_dtype": in_var.dtype,
                             "out_dtype": out_var.dtype,
-                            "op_device": op_device
+                            "op_device": op_device,
+                            "op_role": op.attr("op_role"),
                         })
                     num_cast_ops += 1
                 _rename_arg(op, in_var.name, out_var.name)
@@ -241,7 +242,8 @@ def _insert_cast_post_op(block, op, idx, src_dtype, dest_dtype, target_name,
             attrs={
                 "in_dtype": target_var.dtype,
                 "out_dtype": cast_var.dtype,
-                "op_device": op.attr("op_device")
+                "op_device": op.attr("op_device"),
+                "op_role": op.attr("op_role"),
             })
         num_cast_ops += 1
         op_var_rename_map[block.idx][target_var.name] = cast_var.name
@@ -415,7 +417,9 @@ def cast_model_to_fp16(program, amp_lists=None, use_fp16_guard=True):
                 keep_fp32_ops.add(op)
                 continue  # processed below
             for in_name in op.input_names:
-                if _keep_fp32_input(op, in_name):
+                # for ipu, all inputs must be converted to fp16
+                if not core.is_compiled_with_ipu() and _keep_fp32_input(
+                        op, in_name):
                     continue
                 for in_var_name in op.input(in_name):
                     in_var = None
@@ -443,7 +447,9 @@ def cast_model_to_fp16(program, amp_lists=None, use_fp16_guard=True):
                         format(op.type, in_var_name, in_var.dtype))
 
             for out_name in op.output_names:
-                if _keep_fp32_output(op, out_name):
+                # for ipu, all outputs must be converted to fp16
+                if not core.is_compiled_with_ipu() and _keep_fp32_output(
+                        op, out_name):
                     continue
                 for out_var_name in op.output(out_name):
                     out_var = None

python/paddle/fluid/tests/unittests/ipu/op_test_ipu.py

@@ -16,7 +16,7 @@ import os
 import random
 import unittest
 import numpy as np
-from enum import Enum
+from enum import IntEnum
 
 import paddle
 import paddle.static
@@ -33,17 +33,24 @@ map_np_dtype_to_fluid_dtype = {
 }
 
 
-class ExecutionMode(Enum):
+class ExecutionModeFull(IntEnum):
+    # Run fp32 model on cpu
     CPU_FP32 = 1
+    # Run fp32 model on ipu
     IPU_FP32 = 2
-    # enable_fp16 through ipu_strategy.enable_fp16
+    # Convert model to fp16 using popart transform
+    # All parameters will be converted to fp16
+    # TODO rename to IPU_FP16
     IPU_POPART_FP16 = 3
+    # Mix-precision mode, using `paddle.static.amp.fp16_guard()` to control the
+    # precision of each operator
+    IPU_MIXED_PRECISION = 4
 
-    def __lt__(self, other):
-        return self.value < other.value
 
-    def __gt__(self, other):
-        return self.value > other.value
+class ExecutionMode(IntEnum):
+    CPU_FP32 = ExecutionModeFull.CPU_FP32
+    IPU_FP32 = ExecutionModeFull.IPU_FP32
+    IPU_POPART_FP16 = ExecutionModeFull.IPU_POPART_FP16
 
 
 def np_dtype_to_fluid_str(dtype: np.dtype) -> str:
@@ -61,6 +68,12 @@ class IPUOpTest(unittest.TestCase):
         np.random.seed(cls.SEED)
         random.seed(cls.SEED)
 
+        # For ipu, most ops support fp16
+        cls.amp_list = paddle.static.amp.CustomOpLists(
+            custom_black_list=[], custom_white_list=[])
+        cls.amp_list.unsupported_list = {}
+        cls.amp_list.black_list = {}
+
         # Enable paddle static graph mode
         paddle.enable_static()
 
@@ -114,3 +127,30 @@ class IPUOpTest(unittest.TestCase):
 
             if check_shape:
                 self.assertTrue(ipu_popart_fp16.shape == cpu_fp32.shape)
+
+            ipu_mixed_precision = None
+            if ExecutionModeFull.IPU_MIXED_PRECISION in outputs.keys():
+                ipu_mixed_precision = outputs[
+                    ExecutionModeFull.IPU_MIXED_PRECISION]
+                max_diff = np.abs(
+                    ipu_mixed_precision.astype(np.float32) - cpu_fp32).max()
+                fp16_flag = np.allclose(
+                    ipu_mixed_precision.astype(np.float32),
+                    cpu_fp32,
+                    rtol=self.rtol_fp16,
+                    atol=self.atol_fp16)
+                self.assertTrue(fp16_flag, "max diff is %f" % (max_diff))
+
+                if check_shape:
+                    self.assertTrue(ipu_mixed_precision.shape == cpu_fp32.shape)
+
+            if ExecutionMode.IPU_POPART_FP16 in outputs.keys(
+            ) and ExecutionModeFull.IPU_MIXED_PRECISION in outputs.keys():
+                max_diff = np.abs(ipu_popart_fp16 - ipu_mixed_precision).max()
+                self.assertEqual(ipu_popart_fp16.all(),
+                                 ipu_mixed_precision.all(),
+                                 "max diff is %f" % (max_diff))
+
+                if check_shape:
+                    self.assertTrue(
+                        ipu_popart_fp16.shape == ipu_mixed_precision.shape)

python/paddle/fluid/tests/unittests/ipu/test_mixed_precision_inference_ipu.py

+#  Copyright (c) 2022 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 unittest
+
+import numpy as np
+import paddle
+import paddle.static
+import paddle.nn.functional as F
+from paddle.fluid.tests.unittests.ipu.op_test_ipu import IPUOpTest, ExecutionModeFull
+
+
+@unittest.skipIf(not paddle.is_compiled_with_ipu(),
+                 "core is not compiled with IPU")
+class TestBase(IPUOpTest):
+    def setUp(self):
+        self.set_atol()
+        self.set_data_feed()
+        self.set_feed_attr()
+
+    @property
+    def fp16_enabled(self):
+        return True
+
+    def set_atol(self):
+        self.atol = 1e-6
+        self.rtol = 1e-6
+        self.atol_fp16 = 1e-3
+        self.rtol_fp16 = 1e-3
+
+    def set_data_feed(self):
+        data = np.random.uniform(size=[1, 10, 27, 27])
+        self.feed_fp32 = {"in_0": data.astype(np.float32)}
+        self.feed_fp16 = {"in_0": data.astype(np.float16)}
+
+    def set_feed_attr(self):
+        self.feed_shape = [x.shape for x in self.feed_fp32.values()]
+        self.feed_list = list(self.feed_fp32.keys())
+
+    def dtype_check(self, program, to_fp16_var_names):
+        block = program.global_block()
+        assert len(to_fp16_var_names) > 0
+        for var_name in to_fp16_var_names:
+            assert (block.var(var_name).dtype, paddle.float16)
+
+    def _test_base(self, exec_mode):
+        generator = paddle.fluid.unique_name.UniqueNameGenerator()
+        scope = paddle.static.Scope()
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = self.SEED
+        startup_prog.random_seed = self.SEED
+
+        with paddle.fluid.unique_name.guard(generator):
+            with paddle.static.scope_guard(scope):
+                with paddle.static.program_guard(main_prog, startup_prog):
+                    x = paddle.static.data(
+                        name=self.feed_list[0],
+                        shape=self.feed_shape[0],
+                        dtype='float32')
+
+                    # using fp32
+                    x = paddle.static.nn.conv2d(
+                        input=x, num_filters=3, filter_size=3)
+                    x = paddle.static.nn.batch_norm(x, act='relu')
+                    x = F.max_pool2d(x, kernel_size=2, stride=2)
+
+                    # using fp16
+                    with paddle.static.amp.fp16_guard():
+                        x = paddle.static.nn.conv2d(
+                            input=x, num_filters=6, filter_size=3)
+                        x = paddle.static.nn.batch_norm(x, act='relu')
+                        x = F.max_pool2d(x, kernel_size=2, stride=2)
+
+                    # using fp32
+                    x = paddle.static.nn.fc(x, size=10)
+                    loss = paddle.mean(x)
+                    fetch_list = [loss.name]
+
+                    if exec_mode == ExecutionModeFull.CPU_FP32:
+                        place = paddle.CPUPlace()
+                    else:
+                        place = paddle.IPUPlace()
+
+                    # cast model to fp16
+                    if exec_mode == ExecutionModeFull.IPU_MIXED_PRECISION:
+                        to_fp16_var_names = paddle.static.amp.cast_model_to_fp16(
+                            main_prog, self.amp_list)
+                        self.dtype_check(main_prog, to_fp16_var_names)
+
+                    exe = paddle.static.Executor(place)
+                    exe.run(startup_prog)
+
+                    # cast parameters to fp16
+                    if exec_mode == ExecutionModeFull.IPU_MIXED_PRECISION:
+                        paddle.static.amp.cast_parameters_to_fp16(
+                            paddle.CPUPlace(),
+                            main_prog,
+                            to_fp16_var_names=to_fp16_var_names)
+
+                    if exec_mode != ExecutionModeFull.CPU_FP32:
+                        ipu_strategy = paddle.static.IpuStrategy()
+                        ipu_strategy.set_graph_config(is_training=False)
+                        if exec_mode == ExecutionModeFull.IPU_POPART_FP16:
+                            ipu_strategy.set_precision_config(enable_fp16=True)
+                        program = paddle.static.IpuCompiledProgram(
+                            main_prog, ipu_strategy=ipu_strategy).compile(
+                                self.feed_list, fetch_list)
+                    else:
+                        program = main_prog
+
+                    feed = self.feed_fp32
+                    result = exe.run(program, feed=feed, fetch_list=fetch_list)
+                    return result[0]
+
+    def test(self):
+        output_dict = {}
+        for mode in ExecutionModeFull:
+            if mode == ExecutionModeFull.IPU_POPART_FP16:
+                continue
+            if mode > ExecutionModeFull.IPU_FP32 and not self.fp16_enabled:
+                break
+            output_dict[mode] = self._test_base(mode).flatten()
+
+        self.check(output_dict)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/ipu/test_mixed_precision_training_ipu.py

+#  Copyright (c) 2022 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 unittest
+
+import numpy as np
+import paddle
+import paddle.static
+import paddle.nn.functional as F
+from paddle.fluid.tests.unittests.ipu.op_test_ipu import IPUOpTest, ExecutionModeFull
+
+
+@unittest.skipIf(not paddle.is_compiled_with_ipu(),
+                 "core is not compiled with IPU")
+class TestBase(IPUOpTest):
+    def setUp(self):
+        self.set_atol()
+        self.set_training()
+        self.set_data_feed()
+        self.set_feed_attr()
+
+    @property
+    def fp16_enabled(self):
+        return True
+
+    def set_atol(self):
+        self.atol = 2e-6
+        self.rtol = 1e-5
+        self.atol_fp16 = 1e-2
+        self.rtol_fp16 = 1e-3
+
+    def set_training(self):
+        self.is_training = True
+        self.epoch = 20
+
+    def set_data_feed(self):
+        data = np.random.uniform(size=[1, 3, 28, 28])
+        self.feed_fp32 = {"in_0": data.astype(np.float32)}
+        self.feed_fp16 = {"in_0": data.astype(np.float16)}
+
+    def set_feed_attr(self):
+        self.feed_shape = [x.shape for x in self.feed_fp32.values()]
+        self.feed_list = list(self.feed_fp32.keys())
+
+    def dtype_check(self, program, to_fp16_var_names):
+        block = program.global_block()
+        assert len(to_fp16_var_names) > 0
+        for var_name in to_fp16_var_names:
+            assert (block.var(var_name).dtype, paddle.float16)
+
+    def _test_base(self, exec_mode):
+        generator = paddle.fluid.unique_name.UniqueNameGenerator()
+        scope = paddle.static.Scope()
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = self.SEED
+        startup_prog.random_seed = self.SEED
+
+        with paddle.fluid.unique_name.guard(generator):
+            with paddle.static.scope_guard(scope):
+                with paddle.static.program_guard(main_prog, startup_prog):
+                    x = paddle.static.data(
+                        name=self.feed_list[0],
+                        shape=self.feed_shape[0],
+                        dtype='float32')
+
+                    # using fp32
+                    x = paddle.static.nn.conv2d(
+                        input=x, num_filters=3, filter_size=3)
+                    x = paddle.static.nn.batch_norm(x, act='relu')
+                    x = F.max_pool2d(x, kernel_size=2, stride=2)
+
+                    # using fp16
+                    with paddle.static.amp.fp16_guard():
+                        x = paddle.static.nn.conv2d(
+                            input=x, num_filters=6, filter_size=3)
+                        x = paddle.static.nn.batch_norm(x, act='relu')
+                        x = F.max_pool2d(x, kernel_size=2, stride=2)
+
+                    # using fp32
+                    x = paddle.static.nn.fc(x, size=10)
+                    loss = paddle.mean(x)
+
+                    # optimizer
+                    optimizer = paddle.optimizer.Adam(learning_rate=1e-2)
+                    optimizer.minimize(loss, startup_prog)
+                    fetch_list = [loss.name]
+
+                # cast model to fp16
+                if exec_mode == ExecutionModeFull.IPU_MIXED_PRECISION:
+                    to_fp16_var_names = paddle.static.amp.cast_model_to_fp16(
+                        main_prog, self.amp_list)
+                    self.dtype_check(main_prog, to_fp16_var_names)
+
+                if exec_mode == ExecutionModeFull.CPU_FP32:
+                    place = paddle.CPUPlace()
+                else:
+                    place = paddle.IPUPlace()
+                exe = paddle.static.Executor(place)
+                exe.run(startup_prog)
+
+                # cast parameters to fp16
+                if exec_mode == ExecutionModeFull.IPU_MIXED_PRECISION:
+                    paddle.static.amp.cast_parameters_to_fp16(
+                        paddle.CPUPlace(),
+                        main_prog,
+                        to_fp16_var_names=to_fp16_var_names)
+
+                if exec_mode != ExecutionModeFull.CPU_FP32:
+                    ipu_strategy = paddle.static.IpuStrategy()
+                    ipu_strategy.set_graph_config(is_training=self.is_training)
+                    if exec_mode == ExecutionModeFull.IPU_POPART_FP16:
+                        ipu_strategy.set_precision_config(enable_fp16=True)
+                    program = paddle.static.IpuCompiledProgram(
+                        main_prog, ipu_strategy=ipu_strategy).compile(
+                            self.feed_list, fetch_list)
+                else:
+                    program = main_prog
+
+                feed = self.feed_fp32
+                result = []
+                for i in range(self.epoch):
+                    out = exe.run(program, feed=feed, fetch_list=fetch_list)
+                    result.append(out)
+                return np.array(result)
+
+    def test_base(self):
+        output_dict = {}
+        for mode in ExecutionModeFull:
+            if mode == ExecutionModeFull.IPU_POPART_FP16:
+                continue
+            if mode > ExecutionModeFull.IPU_FP32 and not self.fp16_enabled:
+                break
+            output_dict[mode] = self._test_base(mode).flatten()
+
+        self.check(output_dict)
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/ipu/test_model_parallel_ipu.py

+#  Copyright (c) 2022 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 unittest
+
+import numpy as np
+import paddle
+import paddle.static
+from paddle.fluid.tests.unittests.ipu.op_test_ipu import IPUOpTest
+
+
+@unittest.skipIf(not paddle.is_compiled_with_ipu(),
+                 "core is not compiled with IPU")
+class TestBase(IPUOpTest):
+    def setUp(self):
+        self.set_atol()
+        self.set_training()
+        self.set_attrs()
+        self.set_data_feed()
+
+    def set_training(self):
+        self.is_training = False
+        self.epoch = 10
+
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+
+    def set_data_feed(self):
+        np_image = np.random.rand(1, 3, 10, 10).astype(np.float32)
+        self.feed_cpu = {"image": np_image}
+        self.feed_ipu = {"image": np_image}
+
+    def _test_base(self, run_ipu=True):
+        scope = paddle.static.Scope()
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = self.SEED
+        startup_prog.random_seed = self.SEED
+
+        bs = self.ipu_bs if run_ipu else self.cpu_bs
+        with paddle.static.scope_guard(scope):
+            with paddle.static.program_guard(main_prog, startup_prog):
+                image = paddle.static.data(
+                    name='image', shape=[bs, 3, 10, 10], dtype='float32')
+                with paddle.static.ipu_shard_guard(index=0):
+                    conv1 = paddle.static.nn.conv2d(
+                        image, num_filters=3, filter_size=3, bias_attr=False)
+                with paddle.static.ipu_shard_guard(index=1):
+                    conv2 = paddle.static.nn.conv2d(
+                        conv1, num_filters=3, filter_size=3, bias_attr=False)
+                    # should consider influence of bs
+                    loss = paddle.mean(conv2)
+
+                if self.is_training:
+                    if self.optimizer == 'sgd':
+                        opt = paddle.optimizer.SGD(learning_rate=1e-2)
+                    elif self.optimizer == 'adam':
+                        opt = paddle.optimizer.Adam(learning_rate=1e-2)
+                    elif self.optimizer == 'lamb':
+                        opt = paddle.optimizer.Lamb(learning_rate=1e-2)
+                    else:
+                        raise Exception('optimizer must be sgd, adam or lamb')
+
+                    opt.minimize(loss)
+
+            if run_ipu:
+                place = paddle.IPUPlace()
+            else:
+                place = paddle.CPUPlace()
+            executor = paddle.static.Executor(place)
+            executor.run(startup_prog)
+
+            if run_ipu:
+                feed_list = [image.name]
+                fetch_list = [loss.name]
+                ipu_strategy = paddle.static.IpuStrategy()
+                ipu_strategy.set_graph_config(
+                    num_ipus=2 * self.ipu_options['replicated_graph_count'],
+                    is_training=self.is_training,
+                    enable_manual_shard=True)
+                ipu_strategy.set_options(self.ipu_options)
+                program = paddle.static.IpuCompiledProgram(
+                    main_prog,
+                    ipu_strategy=ipu_strategy).compile(feed_list, fetch_list)
+            else:
+                program = main_prog
+
+            feed = self.feed_ipu if run_ipu else self.feed_cpu
+            epoch = self.epoch
+            if not run_ipu:
+                epoch *= self.ipu_options['replicated_graph_count']
+                epoch *= self.ipu_options['batches_per_step']
+                epoch *= self.ipu_options['accumulation_factor']
+                epoch = epoch / (self.cpu_bs / self.ipu_bs)
+            result = []
+            for i in range(int(epoch)):
+                loss_res = executor.run(program, feed=feed, fetch_list=[loss])
+                result.append(loss_res)
+            return np.array(result).flatten()
+
+    def test(self):
+        cpu_outputs = self._test_base(False)
+        ipu_outputs = self._test_base(True)
+
+        self.assertTrue(np.allclose(cpu_outputs, ipu_outputs, atol=self.atol))
+
+
+class TestReplicaInference(TestBase):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": True,
+            "replicated_graph_count": 2,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+
+    def set_data_feed(self):
+        np_image = np.random.rand(1, 3, 10, 10).astype(np.float32)
+        self.feed_cpu = {"image": np_image}
+        self.feed_ipu = {
+            "image":
+            np.tile(np_image,
+                    [self.ipu_options['replicated_graph_count'], 1, 1, 1])
+        }
+
+
+class TestPipelineInference(TestBase):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 2,
+            "enable_pipelining": True,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+
+    def set_data_feed(self):
+        np_image = np.random.rand(1, 3, 10, 10).astype(np.float32)
+        self.feed_cpu = {"image": np_image}
+        self.feed_ipu = {
+            "image": np.tile(np_image,
+                             [self.ipu_options['batches_per_step'], 1, 1, 1])
+        }
+
+
+class TestTrainBase(TestBase):
+    def set_training(self):
+        self.is_training = True
+        self.epoch = 10
+
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+        self.optimizer = 'sgd'
+
+
+class TestReplicaTrain(TestTrainBase):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": True,
+            "replicated_graph_count": 2,
+        }
+        self.cpu_bs = 2
+        self.ipu_bs = 1
+        self.optimizer = 'sgd'
+
+    def set_data_feed(self):
+        np_image = np.random.rand(1, 3, 10, 10).astype(np.float32)
+        self.feed_cpu = {"image": np.tile(np_image, [self.cpu_bs, 1, 1, 1])}
+        self.feed_ipu = {
+            "image":
+            np.tile(np_image,
+                    [self.ipu_options['replicated_graph_count'], 1, 1, 1])
+        }
+
+    def test(self):
+        cpu_outputs = self._test_base(False)
+        ipu_outputs = self._test_base(True)[::2]
+
+        self.assertTrue(np.allclose(cpu_outputs, ipu_outputs, atol=self.atol))
+
+
+class TestPipelineTrain(TestTrainBase):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 3,
+            "enable_pipelining": True,
+            "enable_gradient_accumulation": True,
+            "accumulation_factor": 3,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 3
+        self.ipu_bs = 1
+        self.optimizer = 'sgd'
+
+    def set_data_feed(self):
+        np_image = np.random.rand(1, 3, 10, 10).astype(np.float32)
+        self.feed_cpu = {"image": np.tile(np_image, [self.cpu_bs, 1, 1, 1])}
+        bps_acc = self.ipu_options['batches_per_step'] * self.ipu_options[
+            'accumulation_factor']
+        self.feed_ipu = {"image": np.tile(np_image, [bps_acc, 1, 1, 1])}
+
+    def test(self):
+        cpu_outputs = self._test_base(False)
+        ipu_outputs = self._test_base(True)[::3]
+
+        self.assertTrue(np.allclose(cpu_outputs, ipu_outputs, atol=self.atol))
+
+
+class TestAdamTrain(TestTrainBase):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+        self.optimizer = 'adam'
+
+
+class TestAdamReplicaTrain(TestReplicaTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": True,
+            "replicated_graph_count": 2,
+        }
+        self.cpu_bs = 2
+        self.ipu_bs = 1
+        self.optimizer = 'adam'
+
+
+class TestAdamPipelineTrain(TestPipelineTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 3,
+            "enable_pipelining": True,
+            "enable_gradient_accumulation": True,
+            "accumulation_factor": 3,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 3
+        self.ipu_bs = 1
+        self.optimizer = 'adam'
+
+
+class TestAdamRecomputationTrain(TestPipelineTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 3,
+            "enable_pipelining": True,
+            "enable_gradient_accumulation": True,
+            "accumulation_factor": 3,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+            "auto_recomputation": 3,
+        }
+        self.cpu_bs = 3
+        self.ipu_bs = 1
+        self.optimizer = 'adam'
+
+
+class TestLambTrain(TestAdamTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 1
+        self.ipu_bs = 1
+        self.optimizer = 'lamb'
+
+
+class TestLambReplicaTrain(TestAdamReplicaTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 1,
+            "enable_pipelining": False,
+            "enable_gradient_accumulation": False,
+            "accumulation_factor": 1,
+            "enable_replicated_graphs": True,
+            "replicated_graph_count": 2,
+        }
+        self.cpu_bs = 2
+        self.ipu_bs = 1
+        self.optimizer = 'lamb'
+
+
+class TestLambPipelineTrain(TestAdamPipelineTrain):
+    def set_attrs(self):
+        self.ipu_options = {
+            "batches_per_step": 3,
+            "enable_pipelining": True,
+            "enable_gradient_accumulation": True,
+            "accumulation_factor": 3,
+            "enable_replicated_graphs": False,
+            "replicated_graph_count": 1,
+        }
+        self.cpu_bs = 3
+        self.ipu_bs = 1
+        self.optimizer = 'lamb'
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/ipu/test_weight_decay_ipu.py

+#  Copyright (c) 2022 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 unittest
+
+import numpy as np
+import paddle
+import paddle.static
+from paddle.fluid.tests.unittests.ipu.op_test_ipu import IPUOpTest
+
+
+@unittest.skipIf(not paddle.is_compiled_with_ipu(),
+                 "core is not compiled with IPU")
+@unittest.skipIf(IPUOpTest.use_ipumodel(), "skip for ipumodel")
+class TestBase(IPUOpTest):
+    def setUp(self):
+        self.set_atol()
+        self.set_data_feed()
+        self.set_feed_attr()
+        self.set_attrs()
+
+    def set_atol(self):
+        self.atol = 1e-6
+
+    def set_data_feed(self):
+        self.feed = {
+            "image": np.random.uniform(size=[1, 3, 10, 10]).astype('float32'),
+        }
+
+    def set_feed_attr(self):
+        self.feed_shape = [x.shape for x in self.feed.values()]
+        self.feed_list = list(self.feed.keys())
+        self.feed_dtype = [x.dtype for x in self.feed.values()]
+
+    def set_attrs(self):
+        self.attrs = {
+            "weight_decay": 4.0,
+            "loss_scaling": 1.0,
+        }
+
+    def _test_optimizer(self, run_ipu=True):
+        def exclude_fn(param):
+            return param.name.endswith('.w_0')
+
+        scope = paddle.static.Scope()
+        main_prog = paddle.static.Program()
+        startup_prog = paddle.static.Program()
+        main_prog.random_seed = self.SEED
+        startup_prog.random_seed = self.SEED
+        np.random.seed(self.SEED)
+
+        with paddle.static.scope_guard(scope):
+            with paddle.static.program_guard(main_prog, startup_prog):
+                image = paddle.static.data(
+                    name='image', shape=[1, 3, 10, 10], dtype='float32')
+                bias = paddle.fluid.layers.create_parameter(
+                    shape=[1, 3, 10, 10], is_bias=True, dtype='float32')
+                add1 = image + bias
+                conv1 = paddle.static.nn.conv2d(
+                    add1, num_filters=3, filter_size=3, bias_attr=False)
+
+                loss = paddle.mean(conv1)
+                opt = paddle.optimizer.Lamb(
+                    learning_rate=1e-1,
+                    lamb_weight_decay=self.attrs['weight_decay'],
+                    exclude_from_weight_decay_fn=exclude_fn)
+                opt.minimize(loss)
+
+            if run_ipu:
+                place = paddle.IPUPlace()
+            else:
+                place = paddle.CPUPlace()
+            exe = paddle.static.Executor(place)
+            exe.run(startup_prog)
+            paddle.static.save(main_prog, "weight_decay")
+
+            if run_ipu:
+                feed_list = [image.name]
+                fetch_list = [loss.name]
+                ipu_strategy = paddle.static.IpuStrategy()
+                ipu_strategy.set_graph_config(is_training=True)
+                ipu_strategy.set_options({
+                    'loss_scaling': self.attrs["loss_scaling"]
+                })
+                program = paddle.static.IpuCompiledProgram(
+                    main_prog, ipu_strategy=ipu_strategy).compile(feed_list,
+                                                                  fetch_list)
+            else:
+                program = main_prog
+
+            result = []
+            for epoch in range(100):
+                loss_res = exe.run(program, feed=self.feed, fetch_list=[loss])
+                result.append(loss_res)
+
+            return np.array(result)
+
+    def test(self):
+        # cpu and ipu dimenstion mismatch, cpu:(100, 1, 1), ipu:(100, 1)
+        ipu_loss = self._test_optimizer(True).flatten()
+        cpu_loss = self._test_optimizer(False).flatten()
+
+        self.assertTrue(np.allclose(ipu_loss, cpu_loss, atol=self.atol))
+
+
+if __name__ == "__main__":
+    unittest.main()