[EP ABI] Add support for creating EP Context models. (microsoft#25124)

### Description
- Updates `OrtEp::Compile()` to allow a plugin EP to create and return
EPContext nodes.
- Updates the example EP plugin to generate an example EPContext model:
<img width="747" alt="image"
src="https://github.com/user-attachments/assets/e5d98a10-ec15-45aa-bfaf-887d3b6226e2"
/>


### Motivation and Context
Adds more of the functionality missing from the EP ABI used for plugin
EPs.

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: Edward Chen <18449977+edgchen1@users.noreply.github.com>

Author: 3 people

cmake/onnxruntime_unittests.cmake

@@ -1834,6 +1834,8 @@ if (WIN32 AND onnxruntime_BUILD_SHARED_LIB AND
     NOT CMAKE_SYSTEM_NAME STREQUAL "Emscripten" AND
     NOT onnxruntime_MINIMAL_BUILD)
   onnxruntime_add_shared_library_module(example_plugin_ep
+                                        ${TEST_SRC_DIR}/autoep/library/example_plugin_ep_utils.h
+                                        ${TEST_SRC_DIR}/autoep/library/example_plugin_ep_utils.cc
                                         ${TEST_SRC_DIR}/autoep/library/example_plugin_ep.cc)
   target_include_directories(example_plugin_ep PRIVATE ${REPO_ROOT}/include/onnxruntime/core/session)
   target_link_libraries(example_plugin_ep PRIVATE onnxruntime)

include/onnxruntime/core/session/onnxruntime_c_api.h

@@ -3678,7 +3678,7 @@ struct OrtApi {
    *
    * \param[in] name Name of the attribute
    * \param[in] data Data content of the attribute
-   * \param[in] len Number of bytes stored in data
+   * \param[in] len Number of elements if data represents an array (e.g., ORT_OP_ATTR_INTS). Otherwise, set to 1.
    * \param[in] type Data type
    * \param[out] op_attr Attribute that has been created, which must be released by OrtApi::ReleaseOpAttr
    *

include/onnxruntime/core/session/onnxruntime_ep_c_api.h

@@ -182,6 +182,14 @@ struct OrtEp {
   /** \brief Compile OrtGraph instances assigned to the OrtEp. Implementer must set a OrtNodeComputeInfo instance
    * for each OrtGraph in order to define its computation function.
    *
+   * If the session is configured to generate a pre-compiled model, the execution provider must return EPContext nodes,
+   * as OrtNode instances, that ONNX Runtime uses to create a pre-compiled model, known as an "EPContext model".
+   * An EPContext model contains EPContext nodes. Each EPContext node encapsulates the pre-compiled binary data for a
+   * OrtGraph compiled for a specific execution provider. For more details about the EPContext design, refer to:
+   *  \htmlonly
+   *  <a href="https://onnxruntime.ai/docs/execution-providers/EP-Context-Design.html">EPContext design document.</a>
+   *  \endhtmlonly
+   *
    * \param[in] this_ptr The OrtEp instance.
    * \param[in] graphs Array of `count` OrtGraph instances to compile. Each graph contains only the nodes for
    *                   which the execution provider indicated support. Nested subgraphs contained by a
@@ -190,9 +198,15 @@ struct OrtEp {
    *                        Each fused node is an OrtNode initialized with the intended fused node name and
    *                        input/output information.
    * \param[in] count The number of OrtGraph instances to compile.
-   * \param[inout] node_compute_infos Array of `count` OrtNodeComputeInfo instances that define each OrtGraph instance's
-   *                                  computation function. The implementer allocates the OrtNodeComputeInfo instances.
-   *                                  ORT calls ReleaseNodeComputeInfos() to release multiple instances in a batch.
+   * \param[out] node_compute_infos Array of `count` OrtNodeComputeInfo instances that define each OrtGraph instance's
+   *                                computation function. The implementer allocates the OrtNodeComputeInfo instances.
+   *                                ORT calls ReleaseNodeComputeInfos() to release multiple instances in a batch.
+   * \param[out] ep_context_nodes Output array of `count` OrtNode instances, each representing an EPContext
+   *                              node for a compiled OrtGraph. The execution provider must use
+   *                              OrtModelEditorApi::CreateNode to create the OrtNode instances. ONNX Runtime takes
+   *                              ownership of the OrtNode instances, so the execution provider must NOT call
+   *                              OrtApi::ReleaseNode. Should be ignored if the session is not configured to generate an
+   *                              EPContext model.
    *
    * \snippet{doc} snippets.dox OrtStatus Return Value
    *
@@ -204,7 +218,8 @@ struct OrtEp {
    */
   OrtStatus*(ORT_API_CALL* Compile)(_In_ OrtEp* this_ptr, _In_ const OrtGraph** graphs,
                                     _In_ const OrtNode** fused_nodes, _In_ size_t count,
-                                    _Out_writes_all_(count) OrtNodeComputeInfo** node_compute_infos);
+                                    _Out_writes_all_(count) OrtNodeComputeInfo** node_compute_infos,
+                                    _Out_writes_(count) OrtNode** ep_context_nodes);
 
   /** \brief Release OrtNodeComputeInfo instances.
    *

onnxruntime/core/session/ep_plugin_provider_interfaces.cc

@@ -8,15 +8,18 @@
 #include <unordered_set>
 #include <utility>
 #include <vector>
+#include "core/framework/abi_pointer_array.h"
 #include "core/framework/compute_capability.h"
 #include "core/framework/error_code_helper.h"
 #include "core/framework/model_metadef_id_generator.h"
 #include "core/graph/ep_api_types.h"
-#include "core/session/ort_apis.h"
+#include "core/graph/model_editor_api_types.h"
 #include "core/session/abi_devices.h"
 #include "core/session/abi_ep_types.h"
 #include "core/session/abi_logger.h"
+#include "core/session/abi_session_options_impl.h"
 #include "core/session/allocator_adapters.h"
+#include "core/session/ort_apis.h"
 #include "core/providers/partitioning_utils.h"
 
 namespace onnxruntime {
@@ -48,7 +51,8 @@ PluginExecutionProviderFactory::CreateProvider(const OrtSessionOptions& session_
     ORT_THROW("Error creating execution provider: ", status.ToString());
   }
 
-  auto ep_wrapper = std::make_unique<PluginExecutionProvider>(UniqueOrtEp(ort_ep, OrtEpDeleter(ep_factory_)));
+  auto ep_wrapper = std::make_unique<PluginExecutionProvider>(UniqueOrtEp(ort_ep, OrtEpDeleter(ep_factory_)),
+                                                              session_options);
   ep_wrapper->SetLogger(session_logger.ToInternal());
 
   return ep_wrapper;
@@ -80,9 +84,10 @@ struct PluginEpMetaDefNameFunctor {
 // PluginExecutionProvider
 //
 
-PluginExecutionProvider::PluginExecutionProvider(UniqueOrtEp ep)
+PluginExecutionProvider::PluginExecutionProvider(UniqueOrtEp ep, const OrtSessionOptions& session_options)
     : IExecutionProvider(ep->GetName(ep.get()), OrtDevice()),  // TODO: What to do about OrtDevice for plugins?
       ort_ep_(std::move(ep)) {
+  generate_ep_ctx_model_ = session_options.value.GetEpContextGenerationOptions().enable;
 }
 
 PluginExecutionProvider::~PluginExecutionProvider() {
@@ -185,6 +190,87 @@ Status PluginExecutionProvider::FusedNodeState::AddFusedNode(const Node& fused_n
   return Status::OK();
 }
 
+/// <summary>
+/// Converts the EPContext nodes provided by the plugin EP (OrtNode instances) to onnxruntime::Node instances.
+/// Note that the EP plugin uses the model editor API to create the OrtNode instances.
+/// </summary>
+/// <param name="ep_name">Name of the plugin EP.</param>
+/// <param name="plugin_ep_context_nodes">EPContext nodes provided by the plugin EP.</param>
+/// <param name="result_nodes">Output parameter set to the resulting array of EPContext nodes.</param>
+/// <param name="result_node_args">Output parameter that stores the NodeArgs used by the EPContext nodes.</param>
+/// <returns>A status indicating success or an error.</returns>
+static Status ConvertEpContextNodes(const std::string& ep_name, const std::vector<OrtNode*> plugin_ep_context_nodes,
+                                    /*out*/ std::vector<std::unique_ptr<Node>>& result_nodes,
+                                    /*out*/ std::vector<std::unique_ptr<NodeArg>>& result_node_args) {
+#if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
+  if (plugin_ep_context_nodes.empty()) {
+    return Status::OK();  // No EPContext nodes.
+  }
+
+  std::vector<std::unique_ptr<Node>> ep_context_nodes_holder;
+  std::vector<std::unique_ptr<NodeArg>> ep_context_node_args_holder;
+
+  ep_context_nodes_holder.reserve(plugin_ep_context_nodes.size());
+
+  for (const OrtNode* ort_node : plugin_ep_context_nodes) {
+    ORT_RETURN_IF_NOT(ort_node != nullptr, ep_name, ": OrtEp::Compile() returned a NULL EPContext node.");
+
+    const ModelEditorNode* editor_node = ModelEditorNode::ToInternal(ort_node);
+    ORT_RETURN_IF_NOT(editor_node != nullptr, ep_name, ": OrtEp::Compile() returned OrtNode objects ",
+                      "that were not created with OrtModelEditorApi.");
+
+    // Create NodeArg for each input/output.
+    std::vector<NodeArg*> input_node_args;
+    std::vector<NodeArg*> output_node_args;
+
+    input_node_args.reserve(editor_node->input_names.size());
+    output_node_args.reserve(editor_node->output_names.size());
+
+    for (const std::string& input_name : editor_node->input_names) {
+      auto node_arg = std::make_unique<NodeArg>(input_name, /*p_arg_type*/ nullptr);  // Graph.Resolve() sets type.
+      input_node_args.push_back(node_arg.get());
+      ep_context_node_args_holder.push_back(std::move(node_arg));
+    }
+
+    for (const std::string& output_name : editor_node->output_names) {
+      auto node_arg = std::make_unique<NodeArg>(output_name, /*p_arg_type*/ nullptr);  // Graph.Resolve() sets type.
+      output_node_args.push_back(node_arg.get());
+      ep_context_node_args_holder.push_back(std::move(node_arg));
+    }
+
+    // Create a name -> attribute map.
+    NodeAttributes attributes;
+    attributes.reserve(editor_node->attributes.size());
+
+    for (const ONNX_NAMESPACE::AttributeProto& attr : editor_node->attributes) {
+      attributes.emplace(attr.name(), attr);
+    }
+
+    // Create Node
+    auto internal_node = std::make_unique<Node>(editor_node->node_name,
+                                                editor_node->operator_name,
+                                                "EPContext node for " + ep_name,
+                                                input_node_args,
+                                                output_node_args,
+                                                &attributes,
+                                                editor_node->domain_name);
+
+    ep_context_nodes_holder.push_back(std::move(internal_node));
+  }
+
+  result_nodes = std::move(ep_context_nodes_holder);
+  result_node_args = std::move(ep_context_node_args_holder);
+
+  return Status::OK();
+#else
+  ORT_UNUSED_PARAMETER(ep_name);
+  ORT_UNUSED_PARAMETER(plugin_ep_context_nodes);
+  ORT_UNUSED_PARAMETER(result_nodes);
+  ORT_UNUSED_PARAMETER(result_node_args);
+  return ORT_MAKE_STATUS(ONNXRUNTIME, NOT_IMPLEMENTED, "Creating EPContext models is not supported in this build");
+#endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
+}
+
 common::Status PluginExecutionProvider::Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
                                                 std::vector<NodeComputeInfo>& node_compute_infos) {
   const logging::Logger* logger = GetLogger();
@@ -220,8 +306,21 @@ common::Status PluginExecutionProvider::Compile(const std::vector<FusedNodeAndGr
     api_fused_nodes.push_back(ep_fused_node->ToExternal());
   }
 
-  ORT_RETURN_IF_ERROR(ToStatusAndRelease(ort_ep_->Compile(ort_ep_.get(), api_graphs.data(), api_fused_nodes.data(),
-                                                          num_graphs, api_node_compute_infos.data())));
+  // Provide an output buffer for the plugin EP to store EPContext nodes if it needs to (i.e., enabled in session options).
+  std::vector<std::unique_ptr<OrtNode, decltype(&OrtApis::ReleaseNode)>> plugin_ep_context_nodes_holder;
+  std::vector<OrtNode*> plugin_ep_context_nodes;
+  plugin_ep_context_nodes_holder.reserve(num_graphs);
+  plugin_ep_context_nodes.resize(num_graphs, nullptr);
+
+  Status compile_status = ToStatusAndRelease(ort_ep_->Compile(ort_ep_.get(), api_graphs.data(), api_fused_nodes.data(),
+                                                              num_graphs, api_node_compute_infos.data(),
+                                                              plugin_ep_context_nodes.data()));
+
+  // Store any EPContext nodes provided by the plugin EP in std::unique_ptr so that they are always properly released.
+  for (OrtNode* ort_node : plugin_ep_context_nodes) {
+    auto unique_ort_node = std::unique_ptr<OrtNode, decltype(&OrtApis::ReleaseNode)>(ort_node, OrtApis::ReleaseNode);
+    plugin_ep_context_nodes_holder.push_back(std::move(unique_ort_node));
+  }
 
   // Save OrtNodeComputeInfo created by OrtEp instance. They're freed when this IExecutionProvider
   // is destroyed.
@@ -231,6 +330,8 @@ common::Status PluginExecutionProvider::Compile(const std::vector<FusedNodeAndGr
     }
   }
 
+  ORT_RETURN_IF_ERROR(compile_status);
+
   // Initialize node_compute_infos as wrappers to api_node_compute_infos.
   for (size_t i = 0; i < num_graphs; i++) {
     OrtNodeComputeInfo* api_node_compute_info = api_node_compute_infos[i];
@@ -268,6 +369,25 @@ common::Status PluginExecutionProvider::Compile(const std::vector<FusedNodeAndGr
     node_compute_infos.push_back(std::move(compute_info));
   }
 
+  // Convert the EPContext nodes provided by the plugin EP into onnxruntime::Node instances.
+  // We store the converted Node and NodeArg instances as members to ensure they can be returned to the ORT graph
+  // partitioner via a call to IExecutionProvider::GetEpContextNodes().
+  if (generate_ep_ctx_model_) {
+    ORT_RETURN_IF_ERROR(ConvertEpContextNodes(Type(), plugin_ep_context_nodes,
+                                              /*out*/ ep_context_nodes_, /*out*/ ep_context_node_args_));
+  }
+
   return Status::OK();
 }
+
+const InlinedVector<const Node*> PluginExecutionProvider::GetEpContextNodes() const {
+  InlinedVector<const Node*> result;
+
+  for (const std::unique_ptr<Node>& node : ep_context_nodes_) {
+    result.push_back(node.get());
+  }
+
+  return result;
+}
+
 }  // namespace onnxruntime

onnxruntime/core/session/ep_plugin_provider_interfaces.h

@@ -17,6 +17,7 @@
 namespace onnxruntime {
 struct EpNode;
 struct EpValueInfo;
+class NodeArg;
 
 /// <summary>
 /// IExecutionProviderFactory that wraps a OrtEpFactory. Required for SessionOptionsAppendExecutionProvider_V2.
@@ -59,7 +60,7 @@ using UniqueOrtEp = std::unique_ptr<OrtEp, OrtEpDeleter>;
 /// </summary>
 class PluginExecutionProvider : public IExecutionProvider {
  public:
-  explicit PluginExecutionProvider(UniqueOrtEp ep);
+  explicit PluginExecutionProvider(UniqueOrtEp ep, const OrtSessionOptions& session_options);
   ~PluginExecutionProvider();
 
   std::vector<std::unique_ptr<ComputeCapability>>
@@ -71,6 +72,8 @@ class PluginExecutionProvider : public IExecutionProvider {
   common::Status Compile(const std::vector<FusedNodeAndGraph>& fused_nodes_and_graphs,
                          std::vector<NodeComputeInfo>& node_compute_funcs) override;
 
+  const InlinedVector<const Node*> GetEpContextNodes() const override;
+
  private:
   struct FusedNodeState {
     FusedNodeState() = default;
@@ -83,12 +86,19 @@ class PluginExecutionProvider : public IExecutionProvider {
   };
 
   UniqueOrtEp ort_ep_;
+  bool generate_ep_ctx_model_ = false;
   std::vector<OrtNodeComputeInfo*> api_node_compute_infos_;
 
   // Fused nodes have to be valid throughout model inference because they may be cached in NodeComputeInfo instances.
   // For each fused node, the Compile() function creates EpNode and EpValueInfo instances on the heap,
   // which are then passed to the underlying OrtEp instance. This class stores this "fused node state"
   // so that it is not destroyed until the EP itself is destroyed.
   std::vector<FusedNodeState> fused_node_states_;
+
+  // Stores the EPContext Nodes created from the OrtNode instances returned by the underlying plugin EP.
+  // Need to store both the Node and NodeArg instances so that they are available when the GraphPartitioner
+  // calls IExecutionProvider::GetEpContextNodes().
+  std::vector<std::unique_ptr<Node>> ep_context_nodes_;
+  std::vector<std::unique_ptr<NodeArg>> ep_context_node_args_;
 };
 }  // namespace onnxruntime

onnxruntime/core/session/provider_policy_context.cc

@@ -292,7 +292,7 @@ Status ProviderPolicyContext::CreateExecutionProvider(const Environment& env, Or
     ORT_RETURN_IF_ERROR(ToStatusAndRelease(info.ep_factory->CreateEp(info.ep_factory, info.devices.data(),
                                                                      info.ep_metadata.data(), info.devices.size(),
                                                                      &options, &logger, &api_ep)));
-    ep = std::make_unique<PluginExecutionProvider>(UniqueOrtEp(api_ep, OrtEpDeleter(*info.ep_factory)));
+    ep = std::make_unique<PluginExecutionProvider>(UniqueOrtEp(api_ep, OrtEpDeleter(*info.ep_factory)), options);
   }
 
   return Status::OK();