Support packed variable for tf data captured function.

If a packed variable is passed to a DatasetOp through a captured function,
1) function instantiation (graph expansion) would insert a PackOp between per-replica arg nodes and the DatasetOp.
2) function execution would feed unpacked ResourceHandles to corresponding sub-functions.

PiperOrigin-RevId: 317752101
Change-Id: I53f7f1d6d075fd40a76a32f329f2dcbbf1db6494

tensorflow/core/common_runtime/process_function_library_runtime.cc

@@ -1356,12 +1356,23 @@ void ProcessFunctionLibraryRuntime::Run(
       // "Index"s of _Arg nodes are unique when all arguments are local Tensors.
       for (const auto& it : comp_data.arg_indices) {
         if (it.sub_index >= 0) {
+          const Tensor& t = args[it.index];
+          if (t.dtype() != DT_RESOURCE) {
             return errors::InvalidArgument("Got unexpected sub_index ",
                                            it.sub_index, " for argument ",
                                            it.index);
           }
+          const auto& handles = t.flat<ResourceHandle>();
+          if (it.sub_index >= handles.size()) {
+            return errors::InvalidArgument(
+                "Sub_index ", it.sub_index, "is out of range [0,",
+                handles.size(), ") for argument ", it.index);
+          }
+          comp_args->args.push_back(Tensor(handles(it.sub_index)));
+        } else {
           comp_args->args.push_back(args[it.index]);
         }
+      }
       return Status::OK();
     };
     return RunMultiDevice(new_opts, handle, rets, cleanup_items,

tensorflow/core/common_runtime/process_function_library_runtime_test.cc

@@ -28,6 +28,7 @@ limitations under the License.
 #include "tensorflow/core/framework/resource_var.h"
 #include "tensorflow/core/framework/tensor_testutil.h"
 #include "tensorflow/core/framework/type_index.h"
+#include "tensorflow/core/framework/types.pb.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/core/status_test_util.h"
 #include "tensorflow/core/lib/core/threadpool.h"
@@ -867,14 +868,29 @@ TEST_F(ProcessFunctionLibraryRuntimeTest, MultiDevice_CompositeDevice) {
   inst_opts.input_resource_dtypes_and_shapes[0] = {
       initial_resource_value0.dtype(), initial_resource_value0.shape()};
 
+  // Packed TensorHandle
+  {
     gtl::InlinedVector<TensorValue, 4> handles;
     handles.push_back(TensorValue(&resource_handle0));
     handles.push_back(TensorValue(&resource_handle1));
     TestFunctionPackedArgs args(0, std::move(handles));
     Tensor ret;
-  TF_CHECK_OK(RunWithPackedArgs("AddVarAcrossDevices", opts, {{"T", DT_FLOAT}},
-                                inst_opts, args, {&ret}));
+    TF_CHECK_OK(RunWithPackedArgs("AddVarAcrossDevices", opts,
+                                  {{"T", DT_FLOAT}}, inst_opts, args, {&ret}));
     test::ExpectTensorEqual<float>(ret, test::AsTensor<float>({40, 60}));
+  }
+
+  // Packed Tensor
+  {
+    Tensor arg(DT_RESOURCE, TensorShape({2}));
+    arg.flat<ResourceHandle>()(0) = resource_handle0.scalar<ResourceHandle>()();
+    arg.flat<ResourceHandle>()(1) = resource_handle1.scalar<ResourceHandle>()();
+
+    Tensor ret;
+    TF_CHECK_OK(Run("AddVarAcrossDevices", opts, {{"T", DT_FLOAT}}, inst_opts,
+                    {arg}, {&ret}));
+    test::ExpectTensorEqual<float>(ret, test::AsTensor<float>({40, 60}));
+  }
 }
 
 TEST_F(ProcessFunctionLibraryRuntimeTest, MultiDevice_ResourceOutput_GPU) {

tensorflow/core/kernels/data/captured_function.cc

@@ -571,6 +571,9 @@ Status CapturedFunction::Instantiate(
   DCHECK(lib->device() != nullptr);
   inst_opts.target = lib->device()->name();
 
+  // Maps from a CompositeDevice name to underlying physical device names.
+  absl::flat_hash_map<string, std::vector<string>> composite_devices;
+
   if (inst_opts.is_multi_device_function) {
     // Compute devices of non-captured inputs.
     //
@@ -596,9 +599,29 @@ Status CapturedFunction::Instantiate(
       const auto& input = captured_inputs_[i];
       DataType dtype = input.dtype();
       if (dtype == DT_RESOURCE) {
-        const ResourceHandle& handle = input.flat<ResourceHandle>()(0);
-        inst_opts.input_devices.push_back(handle.device());
-        const auto& dtypes_and_shapes = handle.dtypes_and_shapes();
+        const auto& handles = input.flat<ResourceHandle>();
+        const ResourceHandle& handle0 = handles(0);
+        string composite_device;
+        auto iter = fdef->arg_attr().find(num_non_captured_inputs + i);
+        if (iter != fdef->arg_attr().end()) {
+          auto arg_attr = iter->second.attr().find("_composite_device");
+          if (arg_attr != iter->second.attr().end()) {
+            composite_device = arg_attr->second.s();
+          }
+        }
+        if (!composite_device.empty()) {
+          if (composite_devices.find(composite_device) ==
+              composite_devices.end()) {
+            for (int i = 0; i < handles.size(); ++i) {
+              composite_devices[composite_device].push_back(
+                  handles(i).device());
+            }
+          }
+          inst_opts.input_devices.push_back(composite_device);
+        } else {
+          inst_opts.input_devices.push_back(handle0.device());
+        }
+        const auto& dtypes_and_shapes = handle0.dtypes_and_shapes();
         // Set dtypes and shapes for resource variable inputs.
         if (!dtypes_and_shapes.empty()) {
           input_resource_variable_dtypes_and_shapes[num_non_captured_inputs +
@@ -613,6 +636,10 @@ Status CapturedFunction::Instantiate(
       }
     }
 
+    for (const auto& it : composite_devices) {
+      inst_opts.composite_devices[it.first] = &it.second;
+    }
+
     for (size_t i = 0; i < fdef->signature().output_arg_size(); ++i) {
       inst_opts.output_devices.push_back(inst_opts.target);
     }

tensorflow/python/distribute/custom_training_loop_input_test.py

@@ -796,10 +796,6 @@ class InputIterationTest(test.TestCase, parameterized.TestCase,
           mode=["eager"]
       ))
   def testMultiDeviceDataCapturedFunction(self, distribution):
-    if getattr(distribution, "_enable_packed_variable_in_eager_mode", False):
-      self.skipTest(
-          "Dataset captured function doesn't support packed tensors yet "
-          "(b/145922293).")
     inputs = constant_op.constant([2., 3.])
     dataset = lambda _: dataset_ops.Dataset.from_tensor_slices(inputs).repeat(5)
     input_iterator = iter(

tensorflow/python/eager/function_test.py

@@ -224,6 +224,13 @@ class FunctionTest(test.TestCase, parameterized.TestCase):
 
       return read0, read1, read2, read3
 
+    arg_attrs = read_var.get_concrete_function().function_def.arg_attr
+    self.assertLen(arg_attrs, 2)
+    self.assertEqual(arg_attrs[0].attr['_composite_device'].s,
+                     compat.as_bytes(packed_var_0.device))
+    self.assertEqual(arg_attrs[1].attr['_composite_device'].s,
+                     compat.as_bytes(packed_var_1.device))
+
     self.assertAllEqual(read_var(), (1 + 5, 2 + 5, 3 + 6, 4 + 6))
 
   def testImplementsAttributeBasic(self):

tensorflow/python/framework/func_graph.py

@@ -646,6 +646,13 @@ class FuncGraph(ops.Graph):
     if capture is None:
       placeholder = _create_substitute_placeholder(
           tensor, name=name, dtype=tensor.dtype, shape=shape)
+      # Record the composite device as an attribute to the placeholder.
+      # This attribute would be propogated into the arg_attr of the FunctionDef.
+      # Currently, a packed eager tensor is always placed on a CompositeDevice.
+      if isinstance(tensor, ops.EagerTensor) and tensor.is_packed:
+        placeholder.op._set_attr(  # pylint: disable=protected-access
+            "_composite_device",
+            attr_value_pb2.AttrValue(s=compat.as_bytes(tensor.device)))
       self.add_capture(tensor, placeholder)
     else:
       placeholder = capture[1]