Runtime API to retrieve attributes (#10144)

Summary:

Scan implementation today which is kind of slow. I dont think people wil regularly be scanning for tons of tensors though so perf is not super sensitive. Can always add a cache later if necessary.

Reviewed By: larryliu0820

Differential Revision: D72802654

Author: JacobSzwejbka

runtime/executor/method.cpp

@@ -1593,6 +1593,38 @@ EValue& Method::mutable_input(size_t i) {
   return mutable_value(get_input_index(i));
 }
 
+Result<executorch::aten::Tensor> Method::get_attribute(
+    executorch::aten::string_view name) {
+  auto flatbuffer_values = serialization_plan_->values();
+  size_t counter = 0;
+
+  for (size_t i = 0; i < flatbuffer_values->size(); ++i) {
+    auto serialization_value = flatbuffer_values->Get(i);
+    if (serialization_value->val_type() ==
+        executorch_flatbuffer::KernelTypes::Tensor) {
+      const auto s_tensor = static_cast<const executorch_flatbuffer::Tensor*>(
+          serialization_value->val());
+      if (s_tensor->extra_tensor_info() != nullptr &&
+          s_tensor->extra_tensor_info()->fully_qualified_name() != nullptr &&
+          executorch::aten::string_view{
+              s_tensor->extra_tensor_info()->fully_qualified_name()->c_str(),
+              s_tensor->extra_tensor_info()->fully_qualified_name()->size()} ==
+              name) {
+        if (!this->values_[counter].isTensor()) {
+          ET_LOG(
+              Error,
+              "Attribute tensor not at the expected location. The .pte is likely malformed. Please file a bug report on https://github.com/pytorch/executorch/issues");
+          return Error::Internal;
+        }
+        return this->values_[counter].toTensor();
+      }
+    }
+    ++counter;
+  }
+
+  return Error::NotFound;
+}
+
 size_t Method::outputs_size() const {
   const auto* outputs = serialization_plan_->outputs();
   return outputs == nullptr ? 0 : outputs->size();

runtime/executor/method.h

@@ -192,6 +192,18 @@ class Method final {
    */
   ET_NODISCARD Error get_inputs(EValue* input_evalues, size_t length);
 
+  /**
+   *
+   * Retrieves the attribute tensor associated with the given name.
+   *
+   * @param[in] name The name of the attribute tensor to retrieve.
+   *
+   * @returns Result containing the attribute tensor on success, non-Ok on
+   * failure.
+   */
+  ET_NODISCARD Result<executorch::aten::Tensor> get_attribute(
+      executorch::aten::string_view name) __attribute__((optimize("no-unroll-loops")));
+
   /**
    * Execute the method.
    *

runtime/executor/method_meta.cpp

@@ -69,9 +69,11 @@ TensorInfo::TensorInfo(
     Span<const int32_t> sizes,
     Span<const uint8_t> dim_order,
     executorch::aten::ScalarType scalar_type,
-    const bool is_memory_planned)
+    const bool is_memory_planned,
+    std::optional<executorch::aten::string_view> name)
     : sizes_(sizes),
       dim_order_(dim_order),
+      name_(name),
       scalar_type_(scalar_type),
       is_memory_planned_(is_memory_planned),
       nbytes_(calculate_nbytes(sizes_, scalar_type_)) {}
@@ -96,6 +98,10 @@ size_t TensorInfo::nbytes() const {
   return nbytes_;
 }
 
+std::optional<executorch::aten::string_view> TensorInfo::name() const {
+  return name_;
+}
+
 MethodMeta::MethodMeta(const executorch_flatbuffer::ExecutionPlan* s_plan)
     : s_plan_(s_plan) {}
 
@@ -150,7 +156,8 @@ Result<TensorInfo> MethodMeta::input_tensor_meta(size_t index) const {
       static_cast<executorch::aten::ScalarType>(tensor_value->scalar_type()),
       tensor_value->allocation_info() != nullptr ||
           tensor_value->data_buffer_idx() !=
-              0); // Count constant returns as memory planned.
+              0 /* is_memory_planned */); // Count constant returns as memory
+                                          // planned.
 }
 
 size_t MethodMeta::num_outputs() const {
@@ -201,7 +208,59 @@ Result<TensorInfo> MethodMeta::output_tensor_meta(size_t index) const {
       static_cast<executorch::aten::ScalarType>(tensor_value->scalar_type()),
       tensor_value->allocation_info() != nullptr ||
           tensor_value->data_buffer_idx() !=
-              0); // Count constant returns as memory planned.
+              0 /* is_memory_planned */); // Count constant returns as memory
+                                          // planned.
+}
+
+size_t MethodMeta::num_attributes() const {
+  size_t counter = 0;
+  auto values = s_plan_->values();
+  for (size_t i = 0; i < values->size(); ++i) {
+    auto value = values->Get(i);
+    if (value->val_type() == executorch_flatbuffer::KernelTypes::Tensor) {
+      auto tensor_value = value->val_as_Tensor();
+      if (tensor_value->extra_tensor_info() != nullptr &&
+          tensor_value->extra_tensor_info()->fully_qualified_name()->c_str() !=
+              nullptr) {
+        ++counter;
+      }
+    }
+  }
+  return counter;
+}
+
+Result<TensorInfo> MethodMeta::attribute_tensor_meta(size_t index) const {
+  size_t counter = 0;
+  auto values = s_plan_->values();
+  for (size_t i = 0; i < values->size(); ++i) {
+    auto value = values->Get(i);
+    if (value->val_type() == executorch_flatbuffer::KernelTypes::Tensor) {
+      auto tensor_value = value->val_as_Tensor();
+      if (tensor_value->extra_tensor_info() != nullptr &&
+          tensor_value->extra_tensor_info()->fully_qualified_name()->c_str() !=
+              nullptr) {
+        if (counter == index) {
+          auto t_name =
+              tensor_value->extra_tensor_info()->fully_qualified_name();
+          // Count constant returns as memory planned
+          return TensorInfo(
+              Span<const int32_t>(
+                  tensor_value->sizes()->data(), tensor_value->sizes()->size()),
+              Span<const uint8_t>(
+                  tensor_value->dim_order()->data(),
+                  tensor_value->dim_order()->size()),
+              static_cast<executorch::aten::ScalarType>(
+                  tensor_value->scalar_type()),
+              tensor_value->allocation_info() != nullptr ||
+                  tensor_value->data_buffer_idx() != 0 /* is_memory_planned */,
+              executorch::aten::string_view{t_name->c_str(), t_name->size()});
+        }
+        ++counter;
+      }
+    }
+  }
+  ET_LOG(Error, "No attribute tensor found at index %zu", index);
+  return Error::InvalidArgument;
 }
 
 size_t MethodMeta::num_memory_planned_buffers() const {

runtime/executor/method_meta.h

@@ -62,6 +62,11 @@ class TensorInfo final {
    */
   size_t nbytes() const;
 
+  /**
+   * Returns the fully qualified name of the Tensor if it has one.
+   */
+  std::optional<executorch::aten::string_view> name() const;
+
  private:
   // Let MethodMeta create TensorInfo.
   friend class MethodMeta;
@@ -70,7 +75,8 @@ class TensorInfo final {
       Span<const int32_t> sizes,
       Span<const uint8_t> dim_order,
       executorch::aten::ScalarType scalar_type,
-      const bool is_memory_planned);
+      const bool is_memory_planned,
+      std::optional<executorch::aten::string_view> name = std::nullopt);
 
   /**
    * The sizes of the tensor.
@@ -88,6 +94,9 @@ class TensorInfo final {
    */
   Span<const uint8_t> dim_order_;
 
+  /// The fully qualified name of the Tensor.
+  std::optional<executorch::aten::string_view> name_;
+
   /// The scalar type of the tensor.
   executorch::aten::ScalarType scalar_type_;
 
@@ -170,6 +179,21 @@ class MethodMeta final {
    */
   Result<TensorInfo> output_tensor_meta(size_t index) const;
 
+  /**
+   * Get the number of attribute tensors in this method.
+   *
+   * @returns The number of attribute tensors.
+   */
+  size_t num_attributes() const __attribute__((optimize("no-unroll-loops")));
+
+  /**
+   * Get metadata about the specified attribute tensor.
+   *
+   * @param[in] index The index of the attribute tensor to look up.
+   * @returns The metadata on success, or an error on failure.
+   */
+  Result<TensorInfo> attribute_tensor_meta(size_t index) const __attribute__((optimize("no-unroll-loops")));
+
   /**
    * Get the number of memory-planned buffers this method requires.
    *

runtime/executor/test/CMakeLists.txt

@@ -27,9 +27,10 @@ add_custom_command(
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleLinearProgram.ptd"
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleMultipleEntry.pte"
          "${CMAKE_CURRENT_BINARY_DIR}/ModuleSimpleTrain.pte"
+         "${CMAKE_CURRENT_BINARY_DIR}/ModuleStateful.pte"
   COMMAND
     python3 -m test.models.export_program --modules
-    "ModuleAdd,ModuleAddHalf,ModuleDynamicCatUnallocatedIO,ModuleIndex,ModuleLinear,ModuleMultipleEntry,ModuleSimpleTrain"
+    "ModuleAdd,ModuleAddHalf,ModuleDynamicCatUnallocatedIO,ModuleIndex,ModuleLinear,ModuleMultipleEntry,ModuleSimpleTrain,ModuleStateful"
     --outdir "${CMAKE_CURRENT_BINARY_DIR}" 2> /dev/null
   COMMAND
     python3 -m test.models.export_program --modules "ModuleLinear"
@@ -51,6 +52,7 @@ add_custom_target(
           "${CMAKE_CURRENT_BINARY_DIR}/ModuleLinearProgram.ptd"
           "${CMAKE_CURRENT_BINARY_DIR}/ModuleMultipleEntry.pte"
           "${CMAKE_CURRENT_BINARY_DIR}/ModuleSimpleTrain.pte"
+          "${CMAKE_CURRENT_BINARY_DIR}/ModuleStateful.pte"
 )
 
 set(test_env
@@ -64,6 +66,7 @@ set(test_env
     "ET_MODULE_LINEAR_DATA_PATH=${CMAKE_CURRENT_BINARY_DIR}/ModuleLinearProgram.ptd"
     "ET_MODULE_MULTI_ENTRY_PATH=${CMAKE_CURRENT_BINARY_DIR}/ModuleMultipleEntry.pte"
     "ET_MODULE_SIMPLE_TRAIN_PATH=${CMAKE_CURRENT_BINARY_DIR}/ModuleSimpleTrain.pte"
+    "ET_MODULE_STATEFUL_PATH=${CMAKE_CURRENT_BINARY_DIR}/ModuleStateful.pte"
 )
 
 et_cxx_test(

runtime/executor/test/method_meta_test.cpp

@@ -26,26 +26,34 @@ using torch::executor::util::FileDataLoader;
 
 class MethodMetaTest : public ::testing::Test {
  protected:
-  void SetUp() override {
-    // Create a loader for the serialized ModuleAdd program.
-    const char* path = std::getenv("ET_MODULE_ADD_PATH");
+  void load_program(const char* path, const char* module_name) {
+    // Create a loader for the serialized program.
     Result<FileDataLoader> loader = FileDataLoader::from(path);
     ASSERT_EQ(loader.error(), Error::Ok);
-    loader_ = std::make_unique<FileDataLoader>(std::move(loader.get()));
+    loaders_.insert(
+        {module_name,
+         std::make_unique<FileDataLoader>(std::move(loader.get()))});
 
     // Use it to load the program.
     Result<Program> program = Program::load(
-        loader_.get(), Program::Verification::InternalConsistency);
+        loaders_[module_name].get(),
+        Program::Verification::InternalConsistency);
     ASSERT_EQ(program.error(), Error::Ok);
-    program_ = std::make_unique<Program>(std::move(program.get()));
+    programs_.insert(
+        {module_name, std::make_unique<Program>(std::move(program.get()))});
+  }
+
+  void SetUp() override {
+    load_program(std::getenv("ET_MODULE_ADD_PATH"), "add");
+    load_program(std::getenv("ET_MODULE_STATEFUL_PATH"), "stateful");
   }
 
  private:
   // Must outlive program_, but tests shouldn't need to touch it.
-  std::unique_ptr<FileDataLoader> loader_;
+  std::unordered_map<std::string, std::unique_ptr<FileDataLoader>> loaders_;
 
  protected:
-  std::unique_ptr<Program> program_;
+  std::unordered_map<std::string, std::unique_ptr<Program>> programs_;
 };
 
 namespace {
@@ -67,7 +75,7 @@ void check_tensor(const TensorInfo& tensor_info) {
 } // namespace
 
 TEST_F(MethodMetaTest, MethodMetaApi) {
-  Result<MethodMeta> method_meta = program_->method_meta("forward");
+  Result<MethodMeta> method_meta = programs_["add"]->method_meta("forward");
   ASSERT_EQ(method_meta.error(), Error::Ok);
 
   // Appropriate amount of inputs
@@ -97,11 +105,12 @@ TEST_F(MethodMetaTest, MethodMetaApi) {
 
   // Missing method fails
   EXPECT_EQ(
-      program_->method_meta("not_a_method").error(), Error::InvalidArgument);
+      programs_["add"]->method_meta("not_a_method").error(),
+      Error::InvalidArgument);
 }
 
 TEST_F(MethodMetaTest, TensorInfoApi) {
-  Result<MethodMeta> method_meta = program_->method_meta("forward");
+  Result<MethodMeta> method_meta = programs_["add"]->method_meta("forward");
   ASSERT_EQ(method_meta.error(), Error::Ok);
 
   // Input 1
@@ -138,3 +147,19 @@ TEST_F(MethodMetaTest, TensorInfoApi) {
   EXPECT_EQ(
       method_meta->output_tensor_meta(-1).error(), Error::InvalidArgument);
 }
+
+TEST_F(MethodMetaTest, MethodMetaAttribute) {
+  Result<MethodMeta> method_meta =
+      programs_["stateful"]->method_meta("forward");
+  ASSERT_EQ(method_meta.error(), Error::Ok);
+
+  ASSERT_EQ(method_meta->num_attributes(), 1);
+  auto state = method_meta->attribute_tensor_meta(0);
+  ASSERT_TRUE(state.ok());
+
+  ASSERT_EQ(state->name(), "state");
+  ASSERT_FALSE(state->is_memory_planned());
+
+  auto bad_access = method_meta->attribute_tensor_meta(1);
+  ASSERT_EQ(bad_access.error(), Error::InvalidArgument);
+}

runtime/executor/test/method_test.cpp

@@ -79,6 +79,7 @@ class MethodTest : public ::testing::Test {
     load_program(
         std::getenv("ET_MODULE_DYNAMIC_CAT_UNALLOCATED_IO_PATH"), "cat");
     load_program(std::getenv("ET_MODULE_LINEAR_PATH"), "linear");
+    load_program(std::getenv("ET_MODULE_STATEFUL_PATH"), "stateful");
     load_program(
         std::getenv("DEPRECATED_ET_MODULE_LINEAR_CONSTANT_BUFFER_PATH"),
         "linear_constant_buffer");
@@ -339,6 +340,31 @@ TEST_F(MethodTest, ProgramDataSeparationTest) {
   ASSERT_EQ(err, Error::Ok);
 }
 
+TEST_F(MethodTest, MethodGetAttributeTest) {
+  ManagedMemoryManager mmm(kDefaultNonConstMemBytes, kDefaultRuntimeMemBytes);
+  Result<Method> method =
+      programs_["stateful"]->load_method("forward", &mmm.get());
+  ASSERT_EQ(method.error(), Error::Ok);
+
+  auto res = method->get_attribute("state");
+  ASSERT_TRUE(res.ok());
+  // expect data to be empty
+  EXPECT_EQ(res->const_data_ptr(), nullptr);
+
+  int32_t data = 0;
+  res->set_data(&data);
+
+  // expect data to be set
+  EXPECT_EQ(res->const_data_ptr(), &data);
+
+  // Can execute the method.
+  Error err = method->execute();
+  ASSERT_EQ(err, Error::Ok);
+
+  // Expect the state to be incremented
+  EXPECT_EQ(res->const_data_ptr<int32_t>()[0], 1);
+}
+
 /*
  * TODO(T161163608): Test is disabled due to a resize bug in tensor_index_out of
  * the portable op lib

runtime/executor/test/targets.bzl

@@ -122,6 +122,7 @@ def define_common_targets(is_fbcode = False):
             "ET_MODULE_LINEAR_PATH": "$(location fbcode//executorch/test/models:exported_programs[ModuleLinear.pte])",
             "ET_MODULE_MULTI_ENTRY_PATH": "$(location fbcode//executorch/test/models:exported_programs[ModuleMultipleEntry.pte])",
             "ET_MODULE_SIMPLE_TRAIN_PATH": "$(location fbcode//executorch/test/models:exported_programs[ModuleSimpleTrain.pte])",
+            "ET_MODULE_STATEFUL_PATH": "$(location fbcode//executorch/test/models:exported_programs[ModuleStateful.pte])",
             "ET_MODULE_LINEAR_PROGRAM_PATH": "$(location fbcode//executorch/test/models:exported_program_and_data[ModuleLinear.pte])",
             "ET_MODULE_LINEAR_DATA_PATH": "$(location fbcode//executorch/test/models:exported_program_and_data[ModuleLinear.ptd])",
         }

test/end2end/exported_module.py

@@ -70,6 +70,7 @@ def export(
         skip_type_promotion: bool = False,
         export_joint_graph: bool = False,
         external_constants: bool = False,
+        export_state_names: bool = False,
     ) -> "ExportedModule":
         """
         Creates a new ExportedModule for the specified module class.
@@ -148,7 +149,9 @@ def return_wrapper():
             for method in methods:
                 method_name_to_dynamic_shapes[method] = trace_dynamic_shapes
 
-        memory_planning_pass = MemoryPlanningPass()
+        memory_planning_pass = MemoryPlanningPass(
+            alloc_mutable_buffers=not export_state_names
+        )
         if hasattr(eager_module, "get_memory_planning_pass"):
             memory_planning_pass = eager_module.get_memory_planning_pass()  # type: ignore[operator]
 
@@ -208,6 +211,7 @@ def __init__(self, method):
                 memory_planning_pass=memory_planning_pass,
                 to_out_var_pass=ToOutVarPass(ignore_to_out_var_failure),
                 external_constants=external_constants,
+                emit_mutable_buffer_names=export_state_names,
             )
         )