webgpu: fix RecurrentState graph capture with shared buffer aliasing

src/models/onnxruntime_api.h

@@ -88,6 +88,25 @@ p_session_->Run(nullptr, input_names, inputs, std::size(inputs), output_names, o
 #define ORT_GENAI_HAS_EXPERIMENTAL_C_API 1
 #endif
 
+// Expose Ort::Experimental::Get_*_Fn accessors used by onnxruntime_inline.h without pulling
+// in onnxruntime_experimental_cxx_api.h (which transitively includes onnxruntime_cxx_api.h
+// and conflicts with genai's vendored Ort wrappers). Only the C typedefs from
+// onnxruntime_experimental_c_api.h (already included above) are needed here.
+#if ORT_GENAI_HAS_EXPERIMENTAL_C_API
+namespace Ort {
+namespace Experimental {
+#define ORT_EXPERIMENTAL_API(VER, RET, NAME, ...)                                      \
+  inline OrtExperimental_##NAME##_SinceV##VER##_Fn Get_##NAME##_SinceV##VER##_Fn(      \
+      const OrtApi* api) {                                                             \
+    return reinterpret_cast<OrtExperimental_##NAME##_SinceV##VER##_Fn>(                \
+        api->GetExperimentalFunction(kOrtExperimental_##NAME##_SinceV##VER##_FnName)); \
+  }
+#include "onnxruntime_experimental_c_api.inc"
+#undef ORT_EXPERIMENTAL_API
+}  // namespace Experimental
+}  // namespace Ort
+#endif
+
 // Single gate for OrtModelPackageApi support: API version 28+ and the experimental header
 // available. The OrtModelPackage* wrappers below and model_package.{h,cpp} key off this.
 #if defined(ORT_API_VERSION) && ORT_API_VERSION >= 28 && ORT_GENAI_HAS_EXPERIMENTAL_C_API

src/models/recurrent_state.cpp

@@ -88,20 +88,15 @@ RecurrentState::RecurrentState(State& state)
 
   const int num_layers = static_cast<int>(layer_indices_.size());
 
-  if (!state_.params_->IsPastPresentShareBufferEnabled(model_.config_->model.type)) {
+  share_buffers_ = state_.params_->IsPastPresentShareBufferEnabled(model_.config_->model.type);
+
+  if (state_.params_->use_graph_capture && !share_buffers_) {
     throw std::runtime_error(
-        "RecurrentState requires past_present_share_buffer=true. "
-        "Set past_present_share_buffer to true in genai_config.json.");
+        "Graph capture requires past/present buffer sharing for models with recurrent state. "
+        "Ensure past_present_share_buffer=true in genai_config.json and num_beams=1 "
+        "(beam search disables buffer sharing).");
   }
 
-  // WebGPU prohibits binding the same buffer as both read-only (input) and
-  // read-write (output) storage in the same compute pass, so it must use
-  // separate past/present buffers with swap. All other EPs share buffers
-  // for stable addresses (required by TRT-RTX graph replay, beneficial elsewhere).
-  // TODO: Remove WebGPU special case once the ORT WebGPU EP adds a
-  // LinearAttention kernel with native past/present buffer sharing support.
-  share_buffers_ = model_.p_device_kvcache_->GetType() != DeviceType::WEBGPU;
-
   if (!share_buffers_) {
     pasts_.resize(num_layers * 2);
   }
@@ -132,8 +127,8 @@ void RecurrentState::Add() {
 
   const int num_layers = static_cast<int>(layer_indices_.size());
   for (int i = 0; i < num_layers * 2; ++i) {
-    // Shared: alias input=output for stable addresses.
-    // WebGPU: separate past/present buffers to avoid aliasing violation.
+    // Shared buffers: alias input=output for stable addresses (required for graph capture).
+    // Separate buffers: use distinct past/present allocations with per-step pointer swap.
     state_.inputs_.push_back(share_buffers_ ? presents_[i].get() : pasts_[i].get());
     state_.input_names_.push_back(input_name_strings_[i].c_str());
     state_.outputs_.push_back(presents_[i].get());

src/models/recurrent_state.h

@@ -30,7 +30,8 @@ struct RecurrentState {
   std::vector<std::unique_ptr<OrtValue>> pasts_;
   std::vector<std::unique_ptr<OrtValue>> presents_;
 
-  // WebGPU cannot alias input/output buffers, so it uses separate past/present\n  // with swap. All other EPs share buffers for stable addresses.
+  // Mirrors past_present_share_buffer config: true means inputs alias outputs (same allocation,
+  // stable handles for graph capture). False uses separate past/present buffers with per-step swap.
   bool share_buffers_{false};
   size_t input_index_{~0U};
   size_t output_index_{~0U};