Release captured graph resources when generator is destroyed

.github/workflows/win-cuda-x64-build.yml

@@ -57,6 +57,9 @@ jobs:
         shell: pwsh
         run: |
           $resp = Invoke-RestMethod "${{ env.ORT_NIGHTLY_REST_API }}"
+          # Hold the CUDA build at ORT 1.26.0 because ORT-GPU 1.27 nupkgs link
+          # the CUDA 13 runtime, while this runner ships CUDA 12.8. Keep this
+          # pin aligned with the USE_CUDA branch in cmake/ortlib.cmake.
           $ORT_NIGHTLY_VERSION = "1.26.0"
           Write-Host "$ORT_NIGHTLY_VERSION"
           "ORT_NIGHTLY_VERSION=$ORT_NIGHTLY_VERSION" | Out-File -FilePath $env:GITHUB_ENV -Append

.pipelines/stages/jobs/steps/integration-pytest-step.yml

@@ -29,10 +29,22 @@ steps:
       python -m pip install --upgrade pip
       python -m pip install -r test\python\requirements.txt
       python -m pip install pytest
-      python -m pip install onnxruntime-gpu==1.26.0
+      # CUDA: onnxruntime-gpu 1.27.0 links libcudart.so.13 (CUDA 13); CI runners
+      # ship CUDA 12.x. Hold CUDA on 1.26.0 until the CI image migrates.
+      # WebGPU: ORT 1.27 ships a MatMulNBitsMlpFusion transformer whose kernel
+      # only exists in plugin EP 0.2.0+. Pair ORT 1.27 with the 0.2.0.dev wheel
+      # from the ORT-Nightly Azure feed (not published to PyPI).
+      if ("${{ parameters.ep }}" -eq "cpu") {
+        python -m pip install onnxruntime-gpu==1.27.0
+      } elseif ("${{ parameters.ep }}" -eq "cuda") {
+        python -m pip install onnxruntime-gpu==1.26.0
+      } else {
+        python -m pip install onnxruntime==1.27.0
+      }
       python -m pip install $wheel.FullName
       if ("${{ parameters.ep }}" -eq "webgpu") {
-        python -m pip install onnxruntime-ep-webgpu==0.1.0
+        python -m pip install onnxruntime-ep-webgpu==0.2.0.dev20260611 `
+          --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/
       }
     displayName: 'Install source-built wheel and test deps'
     workingDirectory: '$(Build.Repository.LocalPath)'
@@ -75,6 +87,9 @@ steps:
           PYTHON_EXE=/opt/python/cp312-cp312/bin/python3.12
           \$PYTHON_EXE -m pip install -r test/python/requirements.txt
           \$PYTHON_EXE -m pip install pytest
+          # onnxruntime-gpu 1.27.0 links libcudart.so.13 (CUDA 13). This CUDA
+          # container image ships CUDA 12.x, so the wheel fails to import.
+          # Hold at 1.26.0 until the image migrates to CUDA 13.
           \$PYTHON_EXE -m pip install onnxruntime-gpu==1.26.0
           \$PYTHON_EXE -m pip install /wheel/*.whl
           nvidia-smi --query-gpu=name,compute_cap,driver_version,memory.total,memory.used,memory.free --format=csv
@@ -113,10 +128,14 @@ steps:
       python -m pip install --upgrade pip
       python -m pip install -r test/python/requirements.txt
       python -m pip install pytest
-      python -m pip install onnxruntime==1.26.0
+      # WebGPU: ORT 1.27 ships a MatMulNBitsMlpFusion transformer whose kernel
+      # only exists in plugin EP 0.2.0+. Pair ORT 1.27 with the 0.2.0.dev wheel
+      # from the ORT-Nightly Azure feed (not published to PyPI).
+      python -m pip install onnxruntime==1.27.0
       python -m pip install ${{ parameters.wheelDir }}/*.whl
       if [ "${{ parameters.ep }}" = "webgpu" ]; then
-        python -m pip install onnxruntime-ep-webgpu==0.1.0
+        python -m pip install onnxruntime-ep-webgpu==0.2.0.dev20260611 \
+          --index-url https://aiinfra.pkgs.visualstudio.com/PublicPackages/_packaging/ORT-Nightly/pypi/simple/
       fi
     displayName: 'Install source-built wheel and test deps'
     workingDirectory: '$(Build.Repository.LocalPath)'

cmake/ortlib.cmake

@@ -80,7 +80,7 @@ if(ORT_HOME)
   endif()
 else()
   # If ORT_HOME is not specified, download the onnxruntime headers and libraries from the nightly feed
-  set(ORT_VERSION "1.26.0")
+  set(ORT_VERSION "1.27.0")
   set(ORT_FEED_ORG_NAME "aiinfra")
   set(ORT_FEED_PROJECT "2692857e-05ef-43b4-ba9c-ccf1c22c437c")
   set(ORT_NIGHTLY_FEED_ID "7982ae20-ed19-4a35-a362-a96ac99897b7")
@@ -89,6 +89,12 @@ else()
     set(ORT_VERSION "1.24.4")
     set(ORT_PACKAGE_NAME "Microsoft.ML.OnnxRuntime.DirectML")
   elseif(USE_CUDA)
+    # ORT-GPU 1.27 nupkgs link the CUDA 13 runtime (libcudart.so.13,
+    # libcublasLt.so.13). CI runners ship CUDA 12.x, so the provider .so
+    # fails to load. SessionReleaseCapturedGraph (the API 27 feature) is
+    # WebGPU-only, so holding CUDA at 1.26 does not lose functionality on
+    # the CUDA path. Drop this override once CI images migrate to CUDA 13.
+    set(ORT_VERSION "1.26.0")
     if(CMAKE_SYSTEM_NAME STREQUAL "Linux")
       set(ORT_PACKAGE_NAME "Microsoft.ML.OnnxRuntime.Gpu.Linux")
     elseif(WIN32)

src/models/model.cpp

@@ -64,7 +64,8 @@ State::State(const GeneratorParams& params, const Model& model)
     std::random_device rd;
     std::mt19937 gen(rd());
     std::uniform_int_distribution<> dis(1, INT_MAX);
-    graph_id_ = std::to_string(dis(gen));
+    graph_id_value_ = dis(gen);
+    graph_id_ = std::to_string(graph_id_value_);
   }
 }
 
@@ -94,6 +95,7 @@ void State::Run(OrtSession& session, bool graph_capture_this_run) {
   DurationTrace trace{"State::Run"};
 
   if (params_->use_graph_capture) {
+    graph_capture_session_ = &session;
     if (graph_capture_this_run) {
       run_options_->AddConfigEntry("gpu_graph_id", graph_id_.c_str());
     } else {
@@ -223,6 +225,25 @@ void State::SetActiveAdapter(Adapters* adapters, const std::string& adapter_name
 }
 
 State::~State() {
+  // Release captured graph resources in the EP.
+  // Note: the EP contract is to no-op when the id was never captured (e.g., when
+  // graph_capture_this_run was always false for this State). The call is wrapped
+  // in try/catch because destructors must not throw -- a throw during unwinding
+  // would call std::terminate.
+#if ORT_API_VERSION >= 27
+  if (graph_capture_session_ && graph_id_value_ > 0) {
+    try {
+      graph_capture_session_->ReleaseCapturedGraph(graph_id_value_);
+    } catch (...) {
+      // Best-effort cleanup; swallow to keep the destructor non-throwing.
+      if (g_log.enabled && g_log.ort_lib) {
+        Log("ort_lib") << "ReleaseCapturedGraph(id=" << graph_id_value_
+                       << ") failed: unknown exception" << std::endl;
+      }
+    }
+  }
+#endif
+
   if (adapters_) {
     for (const auto& adapter_name : adapter_names_) {
       adapters_->ReleaseAdapter(adapter_name);

src/models/model.h

@@ -63,6 +63,11 @@ struct State {
 
  private:
   std::string graph_id_{};
+  int graph_id_value_{0};  // integer form of graph_id_, used to avoid re-parsing in the destructor
+  // Session used for graph capture; not owned. Lifetime invariant: the OrtSession
+  // outlives this State because State is owned by Generator, and Generator is
+  // destroyed before the Model (and its session) that produced it.
+  OrtSession* graph_capture_session_{nullptr};
   std::shared_ptr<Adapters> adapters_;
   ExtraOutputs extra_outputs_;
 };

src/models/onnxruntime_api.h

@@ -809,6 +809,10 @@ struct OrtSession {
 
   void SetEpDynamicOptions(_In_opt_ const char* const* keys, const char* const* values, size_t kv_len);
 
+#if ORT_API_VERSION >= 27
+  void ReleaseCapturedGraph(int graph_annotation_id);  ///< Wraps OrtApi::SessionReleaseCapturedGraph (ORT 1.27+)
+#endif
+
   static void operator delete(void* p) { Ort::api->ReleaseSession(reinterpret_cast<OrtSession*>(p)); }
   Ort::Abstract make_abstract;
 };

src/models/onnxruntime_inline.h

@@ -935,6 +935,12 @@ inline void OrtSession::SetEpDynamicOptions(const char* const* keys, const char*
   Ort::ThrowOnError(Ort::api->SetEpDynamicOptions(this, keys, values, kv_len));
 }
 
+#if ORT_API_VERSION >= 27
+inline void OrtSession::ReleaseCapturedGraph(int graph_annotation_id) {
+  Ort::ThrowOnError(Ort::api->SessionReleaseCapturedGraph(this, graph_annotation_id));
+}
+#endif
+
 inline std::string OrtModelMetadata::GetProducerName() const {
   Ort::StringAllocator string_allocator;
   Ort::ThrowOnError(Ort::api->ModelMetadataGetProducerName(this, &string_allocator, &string_allocator.out));

test/python/cpu/ort/requirements.txt

@@ -1 +1 @@
-onnxruntime==1.26.0
+onnxruntime==1.27.0

test/python/cuda/ort/requirements.txt

@@ -1 +1,4 @@
-onnxruntime-gpu==1.26.0
+# onnxruntime-gpu 1.27.0 links libcudart.so.13 (CUDA 13). CI runners
+# currently ship CUDA 12.x, so the wheel fails to import. Hold the python
+# pin at 1.26.0 until the CI image migrates to CUDA 13.
+onnxruntime-gpu==1.26.0