[Dy2St][AMP] Support local enable the auto_cast

python/paddle/base/framework.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+from __future__ import annotations
+
 import collections
 import copy
 import functools
@@ -26,6 +28,7 @@
 import warnings
 from collections.abc import Iterable
 from types import FunctionType, MethodType
+from typing import TYPE_CHECKING
 
 import numpy as np
 
@@ -39,6 +42,9 @@
 from .variable_index import _getitem_static, _setitem_impl_, _setitem_static
 from .wrapped_decorator import signature_safe_contextmanager, wrap_decorator
 
+if TYPE_CHECKING:
+    from paddle.static.amp.fp16_utils import AmpOptions
+
 __all__ = []
 
 EMPTY_VAR_NAME = core.kEmptyVarName()
@@ -2935,8 +2941,11 @@ def __init__(
             # attr for static graph mode cuda graph
             self._cuda_graph_attr = _current_cuda_graph_mode
 
-            # attr for OP should cast in AMP mode
-            self._should_auto_cast: bool = True
+            # attr for OP AMP mode
+            # using dynamic import to avoid cyclic dependency
+            from paddle.static.amp.fp16_utils import DEFAULT_AMP_OPTIONS
+
+            self._amp_options: AmpOptions = DEFAULT_AMP_OPTIONS
 
             op_maker = core.op_proto_and_checker_maker
 
@@ -3695,24 +3704,24 @@ def dist_attr(self, dist_attr):
         """
         self.desc.dist_attr = dist_attr
 
-    def set_auto_cast(self, auto_cast):
+    def set_amp_options(self, amp_options):
         """
         Set auto cast attribute of this Operator.
 
         Args:
-            auto_cast(bool): True if this Operator should cast in AMP mode.
+            amp_options (AmpOptions): AmpOptions of this Operator.
         """
-        self._should_auto_cast = auto_cast
+        self._amp_options = amp_options
 
     @property
-    def should_auto_cast(self):
+    def amp_options(self):
         """
         Get auto cast attribute of this Operator.
 
         Returns:
-            bool: True if this Operator should cast in AMP mode.
+            bool: AmpOptions of this Operator.
         """
-        return self._should_auto_cast
+        return self._amp_options
 
 
 @signature_safe_contextmanager
@@ -6985,7 +6994,7 @@ def _copy_data_info_from(self, other, pruned_origin_block_id_map=None):
                 if other_var.stop_gradient:
                     var.stop_gradient = True
 
-    def _copy_operator_info_from(self, other: "Program"):
+    def _copy_operator_info_from(self, other: Program):
         """
         Copy the information of Operator information from other program.
 
@@ -7001,7 +7010,7 @@ def _copy_operator_info_from(self, other: "Program"):
             )
         for dst_block, src_block in zip(self.blocks, other.blocks):
             for dst_op, src_op in zip(dst_block.ops, src_block.ops):
-                dst_op.set_auto_cast(src_op.should_auto_cast)
+                dst_op.set_amp_options(src_op.amp_options)
 
     def list_vars(self):
         """

python/paddle/jit/dy2static/convert_operators.py

@@ -15,6 +15,7 @@
 from __future__ import annotations
 
 import re
+import warnings
 from contextlib import contextmanager
 
 import paddle
@@ -823,8 +824,10 @@ def convert_auto_cast(
 ):
     from .program_translator import ProgramTranslator
 
-    if enable:
-        raise NotImplementedError("Does not support local switching on amp now")
+    warnings.warn(
+        "paddle.amp.auto_cast is an experimental features in auto parallel."
+        + "This will take no effect in normal dy2static."
+    )
 
     amp_records = ProgramTranslator.get_instance()._amp_records
     main_program = paddle.static.default_main_program()

python/paddle/static/amp/fp16_utils.py

@@ -53,6 +53,16 @@ class AmpOptions:
     use_promote: bool
 
 
+DEFAULT_AMP_OPTIONS = AmpOptions(
+    enable=True,
+    custom_white_list=None,
+    custom_black_list=None,
+    level='O1',
+    dtype='float16',
+    use_promote=True,
+)
+
+
 def _rename_arg(op, old_name, new_name):
     """
     If an op has old_name input and output, rename these input
@@ -609,28 +619,31 @@ def map_block(block, fn, parent_op=None):
         map_block(sub_block, fn, op)
 
 
-def prepare_op_should_auto_cast(
+def prepare_op_amp_options(
     program: paddle.static.Program,
     amp_records: dict[int, list[tuple[AmpOptions, int, int]]],
+    global_amp_options: AmpOptions,
 ):
-    amp_enable_op_map: dict[paddle.static.Operator, bool] = {}
+    op_amp_options_map: dict[paddle.static.Operator, AmpOptions] = {}
 
     def fill_amp_enable_op_map(block, parent_op):
         block_idx = block.idx
         ops = block.ops
         for op in ops:
-            # The top level should be FP16
-            current_op_amp_options = amp_enable_op_map.get(parent_op, True)
+            # Set the default options to global_amp_options if the op has not parent op.
+            current_op_amp_options = op_amp_options_map.get(
+                parent_op, global_amp_options
+            )
             if block_idx in amp_records:
                 for amp_options, start, end in amp_records[block_idx]:
                     if op.idx in range(start, end):
-                        current_op_amp_options = amp_options.enable
+                        current_op_amp_options = amp_options
                         break
-            amp_enable_op_map[op] = current_op_amp_options
+            op_amp_options_map[op] = current_op_amp_options
 
     map_block(program.global_block(), fill_amp_enable_op_map)
-    for op, enable in amp_enable_op_map.items():
-        op.set_auto_cast(enable)
+    for op, enable in op_amp_options_map.items():
+        op.set_amp_options(enable)
 
 
 def cast_model_to_fp16(

test/dygraph_to_static/test_local_cast.py

@@ -18,7 +18,21 @@
 
 import paddle
 from paddle.jit.dy2static.program_translator import ProgramTranslator
-from paddle.static.amp.fp16_utils import prepare_op_should_auto_cast
+from paddle.static.amp.fp16_utils import (
+    DEFAULT_AMP_OPTIONS,
+    AmpOptions,
+    prepare_op_amp_options,
+)
+
+GLOBAL_ENABLE_AMP_OPTIONS = DEFAULT_AMP_OPTIONS
+GLOBAL_DISABLE_AMP_OPTIONS = AmpOptions(
+    enable=False,
+    custom_black_list=DEFAULT_AMP_OPTIONS.custom_black_list,
+    custom_white_list=DEFAULT_AMP_OPTIONS.custom_white_list,
+    level=DEFAULT_AMP_OPTIONS.level,
+    dtype=DEFAULT_AMP_OPTIONS.dtype,
+    use_promote=DEFAULT_AMP_OPTIONS.use_promote,
+)
 
 
 class LocalAutoCastLayer1(paddle.nn.Layer):
@@ -62,24 +76,47 @@ def forward(self, x):
         return x + 1
 
 
+class LocalAutoCastLayer3(paddle.nn.Layer):
+    def __init__(self):
+        super().__init__()
+        self._fc = paddle.nn.Linear(10, 10)
+
+    @paddle.jit.to_static(full_graph=True)
+    def forward(self, x):
+        with paddle.amp.auto_cast(True):
+            x = x.astype("float32")
+            x = self._fc(x)
+            y = self._fc(x) * 2
+        if x[0][0] > 1:
+            x = x + y
+        else:
+            x = x - y
+            x = x * 2
+
+        return x + 1
+
+
 class TestLocalCast(unittest.TestCase):
     def get_auto_cast_ops_info_from_program(self, program):
         auto_cast_ops_info = []
         for block in program.blocks:
             current_block_should_auto_cast = []
             auto_cast_ops_info.append(current_block_should_auto_cast)
             for op in block.ops:
-                current_block_should_auto_cast.append(op.should_auto_cast)
+                current_block_should_auto_cast.append(op.amp_options.enable)
         return auto_cast_ops_info
 
-    def should_auto_cast_for_each_ops(self, layer, input):
+    def should_auto_cast_for_each_ops(self, layer, input, global_amp_options):
         concrete_program, _ = layer.forward.get_concrete_program(input)
         program = concrete_program.main_program
-        prepare_op_should_auto_cast(
-            program, ProgramTranslator.get_instance()._amp_records
+        prepare_op_amp_options(
+            program,
+            ProgramTranslator.get_instance()._amp_records,
+            global_amp_options,
         )
         auto_cast_ops_info = self.get_auto_cast_ops_info_from_program(program)
         paddle.enable_static()
+        # Ensure the cloned program has the same auto_cast ops info
         cloned_program = program.clone()
         paddle.disable_static()
         cloned_auto_cast_ops_info = self.get_auto_cast_ops_info_from_program(
@@ -103,7 +140,9 @@ def test_should_auto_cast_1(self):
             # All else branch in auto_cast(False) block
             [False, False, False],
         ]  # fmt: skip
-        actual = self.should_auto_cast_for_each_ops(layer, input)
+        actual = self.should_auto_cast_for_each_ops(
+            layer, input, GLOBAL_ENABLE_AMP_OPTIONS
+        )
         self.assertEqual(expected, actual)
 
     def test_should_auto_cast_2(self):
@@ -120,7 +159,29 @@ def test_should_auto_cast_2(self):
             # All else branch out of auto_cast(False) block
             [True, True, True],
         ]  # fmt: skip
-        actual = self.should_auto_cast_for_each_ops(layer, input)
+        actual = self.should_auto_cast_for_each_ops(
+            layer, input, GLOBAL_ENABLE_AMP_OPTIONS
+        )
+        self.assertEqual(expected, actual)
+
+    def test_should_auto_cast_3(self):
+        layer = LocalAutoCastLayer3()
+        input = paddle.randn([10, 10])
+        expected = [
+            # There are part of ops in auto_cast(True) block
+            [
+                True, True, True, True, True, True,
+                False, False, False, False, False, False, False, False, False, False,
+            ],
+            # All if branch out of auto_cast(True) block
+            [False, False],
+            # All else branch out of auto_cast(True) block
+            [False, False, False],
+        ]  # fmt: skip
+        actual = self.should_auto_cast_for_each_ops(
+            layer, input, GLOBAL_DISABLE_AMP_OPTIONS
+        )
+
         self.assertEqual(expected, actual)