Extend AutoGraph to functions called by Distribution Strategy's experimental_run_v2.

PiperOrigin-RevId: 253722908

tensorflow/python/autograph/impl/api.py

@@ -201,7 +201,6 @@ def convert(recursive=False, optional_features=None, force_conversion=True):
   def decorator(f):
     """Decorator implementation."""
 
-    @functools.wraps(f)
     def wrapper(*args, **kwargs):
       """Wrapper that calls the converted version of f."""
       with ag_ctx.ControlStatusCtx(
@@ -220,12 +219,15 @@ def convert(recursive=False, optional_features=None, force_conversion=True):
           else:
             raise
 
-    wrapper = tf_decorator.make_decorator(f, wrapper)
+    if inspect.isfunction(f) or inspect.ismethod(f):
+      wrapper = functools.update_wrapper(wrapper, f)
+
+    decorated_wrapper = tf_decorator.make_decorator(f, wrapper)
 
     # Sometimes the decorator is just desugared, making it impossible to detect.
     # This attribute makes detection easier.
-    setattr(wrapper, '__ag_compiled', True)
-    return wrapper
+    setattr(decorated_wrapper, '__ag_compiled', True)
+    return decorated_wrapper
 
   return decorator
 

tensorflow/python/distribute/custom_training_loop_test.py

@@ -97,6 +97,32 @@ class InputIterationTest(test.TestCase, parameterized.TestCase):
       results.append(output)
     self._validate_outputs(results)
 
+  @combinations.generate(
+      combinations.combine(
+          distribution=strategy_combinations.strategies_minus_tpu +
+          [strategy_combinations.tpu_strategy_one_step],
+          mode=["eager"]
+      ))
+  def testRunInFunctionAutoGraphApplication(self, distribution):
+    dataset = self._get_dataset()
+
+    def train_step(data):
+      if math_ops.reduce_sum(data) < 0:
+        return -data
+      return data
+
+    @def_function.function
+    def f_train_step(input_data):
+      return distribution.experimental_local_results(
+          distribution.experimental_run_v2(train_step, args=(input_data,)))
+
+    dist_dataset = distribution.experimental_distribute_dataset(dataset)
+    results = []
+    for x in dist_dataset:
+      output = f_train_step(x)
+      results.append(output)
+    self._validate_outputs(results)
+
   @combinations.generate(
       combinations.combine(
           distribution=strategy_combinations.strategies_minus_tpu,

tensorflow/python/distribute/distribute_lib.py

@@ -101,6 +101,8 @@ import threading
 import weakref
 import six
 
+from tensorflow.python.autograph.core import ag_ctx
+from tensorflow.python.autograph.impl import api as autograph
 from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.distribute import device_util
 from tensorflow.python.distribute import distribution_strategy_context
@@ -716,6 +718,7 @@ class Strategy(object):
       (for example, if running on a single replica).
     """
     with self.scope():
+      fn = autograph.tf_convert(fn, ag_ctx.control_status_ctx())
       return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
 
   def reduce(self, reduce_op, value, axis):

tensorflow/python/distribute/tpu_strategy.py

@@ -25,6 +25,8 @@ import weakref
 
 import numpy as np
 
+from tensorflow.python.autograph.core import ag_ctx
+from tensorflow.python.autograph.impl import api as autograph
 from tensorflow.python.distribute import cross_device_ops as cross_device_ops_lib
 from tensorflow.python.distribute import device_util
 from tensorflow.python.distribute import distribute_lib
@@ -160,10 +162,10 @@ class TPUStrategy(distribute_lib.Strategy):
   # This implementation runs a single step. It does not use infeed or outfeed.
   def experimental_run_v2(self, fn, args=(), kwargs=None):
     """See base class."""
+    fn = autograph.tf_convert(fn, ag_ctx.control_status_ctx())
     return self.extended.tpu_run(fn, args, kwargs)
 
 
-
 @tf_export(v1=["distribute.experimental.TPUStrategy"])
 class TPUStrategyV1(distribute_lib.StrategyV1):
   """TPU distribution strategy implementation."""
@@ -199,6 +201,7 @@ class TPUStrategyV1(distribute_lib.StrategyV1):
   # This implementation runs a single step. It does not use infeed or outfeed.
   def experimental_run_v2(self, fn, args=(), kwargs=None):
     """See base class."""
+    fn = autograph.tf_convert(fn, ag_ctx.control_status_ctx())
     return self.extended.tpu_run(fn, args, kwargs)