com.github.asus4.onnxruntime 0.2.0 → 0.2.2

package/Plugins/iOS~/onnxruntime.xcframework/Info.plist

@@ -6,7 +6,7 @@
 	<array>
 		<dict>
 			<key>LibraryIdentifier</key>
-			<string>macos-arm64_x86_64</string>
+			<string>ios-arm64_x86_64-simulator</string>
 			<key>LibraryPath</key>
 			<string>onnxruntime.framework</string>
 			<key>SupportedArchitectures</key>
@@ -15,34 +15,34 @@
 				<string>x86_64</string>
 			</array>
 			<key>SupportedPlatform</key>
-			<string>macos</string>
+			<string>ios</string>
+			<key>SupportedPlatformVariant</key>
+			<string>simulator</string>
 		</dict>
 		<dict>
 			<key>LibraryIdentifier</key>
-			<string>ios-arm64_x86_64-simulator</string>
+			<string>ios-arm64</string>
 			<key>LibraryPath</key>
 			<string>onnxruntime.framework</string>
 			<key>SupportedArchitectures</key>
 			<array>
 				<string>arm64</string>
-				<string>x86_64</string>
 			</array>
 			<key>SupportedPlatform</key>
 			<string>ios</string>
-			<key>SupportedPlatformVariant</key>
-			<string>simulator</string>
 		</dict>
 		<dict>
 			<key>LibraryIdentifier</key>
-			<string>ios-arm64</string>
+			<string>macos-arm64_x86_64</string>
 			<key>LibraryPath</key>
 			<string>onnxruntime.framework</string>
 			<key>SupportedArchitectures</key>
 			<array>
 				<string>arm64</string>
+				<string>x86_64</string>
 			</array>
 			<key>SupportedPlatform</key>
-			<string>ios</string>
+			<string>macos</string>
 		</dict>
 	</array>
 	<key>CFBundlePackageType</key>

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/coreml_provider_factory.h

@@ -31,9 +31,14 @@ enum COREMLFlags {
   // Create an MLProgram. By default it will create a NeuralNetwork model. Requires Core ML 5 or later.
   COREML_FLAG_CREATE_MLPROGRAM = 0x010,
 
+  // Exclude ANE as sometimes this decrease performance
+  // https://developer.apple.com/documentation/coreml/mlcomputeunits?language=objc
+  // there are four compute units:
+  // MLComputeUnitsCPUAndNeuralEngine|MLComputeUnitsCPUAndGPU|MLComputeUnitsCPUOnly|MLComputeUnitsAll
+  COREML_FLAG_USE_CPU_AND_GPU = 0x020,
   // Keep COREML_FLAG_LAST at the end of the enum definition
   // And assign the last COREMLFlag to it
-  COREML_FLAG_LAST = COREML_FLAG_CREATE_MLPROGRAM,
+  COREML_FLAG_LAST = COREML_FLAG_USE_CPU_AND_GPU,
 };
 
 #ifdef __cplusplus

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/onnxruntime_c_api.h

@@ -38,7 +38,7 @@
  *
  * This value is used by some API functions to behave as this version of the header expects.
  */
-#define ORT_API_VERSION 19
+#define ORT_API_VERSION 20
 
 #ifdef __cplusplus
 extern "C" {
@@ -304,6 +304,7 @@ ORT_RUNTIME_CLASS(Op);
 ORT_RUNTIME_CLASS(OpAttr);
 ORT_RUNTIME_CLASS(Logger);
 ORT_RUNTIME_CLASS(ShapeInferContext);
+ORT_RUNTIME_CLASS(LoraAdapter);
 
 #ifdef _WIN32
 typedef _Return_type_success_(return == 0) OrtStatus* OrtStatusPtr;
@@ -621,6 +622,7 @@ typedef struct OrtMIGraphXProviderOptions {
   const char* migraphx_save_model_path;              // migraphx model path name
   int migraphx_load_compiled_model;                  // migraphx int8 cal table. Default 0 = false, noznero = true
   const char* migraphx_load_model_path;              // migraphx model path name
+  bool migraphx_exhaustive_tune;                     // migraphx tuned compile  Default = false
 } OrtMIGraphXProviderOptions;
 
 /** \brief OpenVINO Provider Options
@@ -643,7 +645,7 @@ typedef struct OrtOpenVINOProviderOptions {
    * Valid settings are one of: "CPU_FP32", "CPU_FP16", "GPU_FP32", "GPU_FP16"
    */
   const char* device_type;
-  unsigned char enable_npu_fast_compile;  ///< 0 = disabled, nonzero = enabled
+  unsigned char enable_npu_fast_compile;
   const char* device_id;
   size_t num_of_threads;  ///< 0 = Use default number of threads
   const char* cache_dir;  // path is set to empty by default
@@ -3649,10 +3651,17 @@ struct OrtApi {
    *     - "73"
    *     - "75"
    *   "device_id": The ID of the device to use when setting 'htp_arch'. Defaults to "0" (for single device).
-       "enable_htp_fp16_precision": Only used for float32 model.
-       Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision.
-         - "0": Default. With fp32 precision.
-         - "1": With fp16 precision.
+   *   "enable_htp_fp16_precision": Used for float32 model for HTP backend.
+   *   Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision.
+   *     - "0": With fp32 precision.
+   *     - "1": Default. With fp16 precision.
+   *   "enable_htp_weight_sharing": Enable QNN weight sharing feature while compiling multiple graphs into one QNN context.
+   *     - "0": Default. Disabled.
+   *     - "1": Enabled.
+   *   "offload_graph_io_quantization": Offload graph input quantization and graph output dequantization to another
+   *   execution provider (typically CPU EP).
+   *     - "0": Default. Disabled. QNN EP will handle quantization and dequantization of graph I/O.
+   *     - "1": Enabled.
    *
    * SNPE supported keys:
    *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
@@ -3778,7 +3787,7 @@ struct OrtApi {
 
   /** \brief Release an OrtCANNProviderOptions
    *
-   * \param[in] the pointer of OrtCANNProviderOptions which will been deleted
+   * \param[in] input The pointer of OrtCANNProviderOptions which will been deleted
    *
    * \since Version 1.13.
    */
@@ -4666,6 +4675,82 @@ struct OrtApi {
                   _In_reads_(num_external_initializer_files) char* const* external_initializer_file_buffer_array,
                   _In_reads_(num_external_initializer_files) const size_t* external_initializer_file_lengths,
                   size_t num_external_initializer_files);
+
+  /** \brief Create an OrtLoraAdapter
+   *
+   * The function attempts to locate file specified by adapter_file_path, read it and create an OrtLoraAdapter
+   * instance. The adapter_file_path should be a valid path to a file that contains a valid Lora Adapter
+   * format. The function attempts to validate the format at load time. The file will always be memory mapped, unless
+   * the platform does not support memory mapping, in which case the file will be read into memory.
+   *
+   * \param[in] adapter_file_path adapter file path.
+   * \param[in] allocator optional pointer to a device allocator. If specified
+   *            data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU.
+   *            The data would still be copied to device if required by the model at inference time.
+   * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with
+   *                  OrtApi::ReleaseLoraAdapter.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(CreateLoraAdapter, const ORTCHAR_T* adapter_file_path, _In_ OrtAllocator* allocator,
+                  _Outptr_ OrtLoraAdapter** out);
+
+  /** \brief Create an OrtLoraAdapter
+   *
+   * The function copies the bytes from the array and creates an OrtLoraAdapter instance.
+   *
+   *
+   * \param[in] bytes pointer to a valid Lora Adapter format buffer.
+   * \param[in] num_bytes length of bytes buffer.
+   * \param[in] allocator optional pointer to a device allocator. If specified
+   *            data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU.
+   *            The data would still be copied to device if required by the model at inference time.
+   * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with
+   *                  OrtApi::ReleaseLoraAdapter.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(CreateLoraAdapterFromArray, _In_ const void* bytes, size_t num_bytes, _In_ OrtAllocator* allocator,
+                  _Outptr_ OrtLoraAdapter** out);
+
+  /** \brief Release an ::OrtLoraAdapter obtained from OrtApi::CreateLoraAdapter
+   */
+  ORT_CLASS_RELEASE(LoraAdapter);
+
+  /** \brief Add the Lora Adapter to the list of active adapters.
+   *
+   * The function adds the Lora Adapter to the list of active adapters. The Lora Adapter must be created with
+   * OrtApi::CreateLoraAdapter or FromArray. The Lora Adapter will be used by the session to run the model.
+   * The instance of the OrtRunOptions can then be used to customize the Run() calls.
+   * More than one OrtLoraAdapter can be active at the same time. Lora Parameters that belong to different
+   * Lora adapters that will be active at the same time must not overlap.
+   * This setting does not affect RunWithBinding.
+   *
+   * \param[in] options OrtRunOptions instance
+   * \param[in] adapter OrtLoraAdapter instance
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(RunOptionsAddActiveLoraAdapter, _Inout_ OrtRunOptions* options, _In_ const OrtLoraAdapter* adapter);
+
+  /// @}
+  /// \name OrtEpDynamicOptions
+  /// @{
+
+  /** \brief Set DynamicOptions for EPs (Execution Providers)
+   *
+   * Valid options can be found in `include\onnxruntime\core\session\onnxruntime_session_options_config_keys.h`
+   * Look for `kOrtEpDynamicOptions`
+   *
+   * \param[in] sess OrtSession
+   * \param[in] keys Array of null terminated UTF8 encoded strings of EP dynamic option keys
+   * \param[in] values Array of null terminated UTF8 encoded string of EP dynamic option values
+   * \param[in] kv_len Number of elements in the keys and values arrays
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(SetEpDynamicOptions, _Inout_ OrtSession* sess, _In_reads_(kv_len) const char* const* keys,
+                  _In_reads_(kv_len) const char* const* values, _In_ size_t kv_len);
 };
 
 /*

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/onnxruntime_cxx_api.h

@@ -508,6 +508,7 @@ ORT_DEFINE_RELEASE(CustomOpDomain);
 ORT_DEFINE_RELEASE(ThreadingOptions);
 ORT_DEFINE_RELEASE(Env);
 ORT_DEFINE_RELEASE(RunOptions);
+ORT_DEFINE_RELEASE(LoraAdapter);
 ORT_DEFINE_RELEASE(Session);
 ORT_DEFINE_RELEASE(SessionOptions);
 ORT_DEFINE_RELEASE(TensorTypeAndShapeInfo);
@@ -736,6 +737,32 @@ struct CustomOpDomain : detail::Base<OrtCustomOpDomain> {
   void Add(const OrtCustomOp* op);  ///< Wraps CustomOpDomain_Add
 };
 
+/// \brief LoraAdapter holds a set of Lora Parameters loaded from a single file
+struct LoraAdapter : detail::Base<OrtLoraAdapter> {
+  using Base = detail::Base<OrtLoraAdapter>;
+  using Base::Base;
+
+  explicit LoraAdapter(std::nullptr_t) {}  ///< Create an empty LoraAdapter object, must be assigned a valid one to be used
+  /// \brief Wraps OrtApi::CreateLoraAdapter
+  ///
+  /// The function attempts to load the adapter from the specified file
+  /// \param adapter_path The path to the Lora adapter
+  /// \param allocator optional pointer to a device allocator. If nullptr, the data stays on CPU. It would still
+  ///        be copied to device if required by the model at inference time.
+  static LoraAdapter CreateLoraAdapter(const std::basic_string<ORTCHAR_T>& adapter_path,
+                                       OrtAllocator* allocator);
+
+  /// \brief Wraps OrtApi::CreateLoraAdapterFromArray
+  ///
+  /// The function attempts to load the adapter from the specified byte array.
+  /// \param bytes The byte array containing file LoraAdapter format
+  /// \param num_bytes The number of bytes in the byte array
+  /// \param allocator optional pointer to a device allocator. If nullptr, the data stays on CPU. It would still
+  ///        be copied to device if required by the model at inference time.
+  static LoraAdapter CreateLoraAdapterFromArray(const void* bytes, size_t num_bytes,
+                                                OrtAllocator* allocator);
+};
+
 /** \brief RunOptions
  *
  */
@@ -766,6 +793,14 @@ struct RunOptions : detail::Base<OrtRunOptions> {
    * Wraps OrtApi::RunOptionsUnsetTerminate
    */
   RunOptions& UnsetTerminate();
+
+  /** \brief Add the LoraAdapter to the list of active adapters.
+   *  The setting does not affect RunWithBinding() calls.
+   *
+   * Wraps OrtApi::RunOptionsAddActiveLoraAdapter
+   * \param adapter The LoraAdapter to be used as the active adapter
+   */
+  RunOptions& AddActiveLoraAdapter(const LoraAdapter& adapter);
 };
 
 namespace detail {
@@ -1105,6 +1140,19 @@ struct SessionImpl : ConstSessionImpl<T> {
    *  The OrtAllocator instances must be valid at the point of memory release.
    */
   AllocatedStringPtr EndProfilingAllocated(OrtAllocator* allocator);  ///< Wraps OrtApi::SessionEndProfiling
+
+  /** \brief Set DynamicOptions for EPs (Execution Providers)
+   *
+   * Wraps OrtApi::SetEpDynamicOptions
+   *
+   * Valid options can be found in `include\onnxruntime\core\session\onnxruntime_session_options_config_keys.h`
+   * Look for `kOrtEpDynamicOptions`
+   *
+   * \param[in] keys Array of null terminated UTF8 encoded strings of EP dynamic option keys
+   * \param[in] values Array of null terminated UTF8 encoded string of EP dynamic option values
+   * \param[in] kv_len Number of elements in the keys and values arrays
+   */
+  void SetEpDynamicOptions(const char* const* keys, const char* const* values, size_t kv_len);
 };
 
 }  // namespace detail

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/onnxruntime_cxx_inline.h

@@ -557,6 +557,20 @@ inline void CustomOpDomain::Add(const OrtCustomOp* op) {
   ThrowOnError(GetApi().CustomOpDomain_Add(p_, op));
 }
 
+inline LoraAdapter LoraAdapter::CreateLoraAdapter(const std::basic_string<ORTCHAR_T>& adapter_path,
+                                                  OrtAllocator* allocator) {
+  OrtLoraAdapter* p;
+  ThrowOnError(GetApi().CreateLoraAdapter(adapter_path.c_str(), allocator, &p));
+  return LoraAdapter{p};
+}
+
+inline LoraAdapter LoraAdapter::CreateLoraAdapterFromArray(const void* bytes, size_t num_bytes,
+                                                           OrtAllocator* allocator) {
+  OrtLoraAdapter* p;
+  ThrowOnError(GetApi().CreateLoraAdapterFromArray(bytes, num_bytes, allocator, &p));
+  return LoraAdapter{p};
+}
+
 inline RunOptions::RunOptions() {
   ThrowOnError(GetApi().CreateRunOptions(&p_));
 }
@@ -609,6 +623,11 @@ inline RunOptions& RunOptions::UnsetTerminate() {
   return *this;
 }
 
+inline RunOptions& RunOptions::AddActiveLoraAdapter(const LoraAdapter& adapter) {
+  ThrowOnError(GetApi().RunOptionsAddActiveLoraAdapter(p_, adapter));
+  return *this;
+}
+
 namespace detail {
 
 template <typename T>
@@ -1074,6 +1093,11 @@ inline AllocatedStringPtr SessionImpl<T>::EndProfilingAllocated(OrtAllocator* al
   return AllocatedStringPtr(out, detail::AllocatedFree(allocator));
 }
 
+template <typename T>
+inline void SessionImpl<T>::SetEpDynamicOptions(const char* const* keys, const char* const* values, size_t kv_len) {
+  ThrowOnError(GetApi().SetEpDynamicOptions(this->p_, keys, values, kv_len));
+}
+
 }  // namespace detail
 
 inline SessionOptions::SessionOptions() {
@@ -2044,6 +2068,9 @@ inline ShapeInferContext::Ints ShapeInferContext::GetAttrInts(const char* attr_n
   int64_t i = {};
   size_t out = {};
   // first call to get the bytes needed
+  // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure.
+  // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success).
+  // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}.
   auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, &i, sizeof(i), &out);
   if (status) {
     size_t num_i = out / sizeof(int64_t);
@@ -2051,6 +2078,9 @@ inline ShapeInferContext::Ints ShapeInferContext::GetAttrInts(const char* attr_n
     Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_INTS, ints.data(), out, &out));
     return ints;
   } else {
+    if (out == 0u) {
+      return {};
+    }
     return {i};
   }
 }
@@ -2068,6 +2098,9 @@ inline ShapeInferContext::Floats ShapeInferContext::GetAttrFloats(const char* at
   float f = {};
   size_t out = {};
   // first call to get the bytes needed
+  // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure.
+  // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success).
+  // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}.
   auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, &f, sizeof(f), &out);
   if (status) {
     size_t num_f = out / sizeof(float);
@@ -2075,6 +2108,9 @@ inline ShapeInferContext::Floats ShapeInferContext::GetAttrFloats(const char* at
     Ort::ThrowOnError(ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_FLOATS, floats.data(), out, &out));
     return floats;
   } else {
+    if (out == 0u) {
+      return {};
+    }
     return {f};
   }
 }
@@ -2099,6 +2135,9 @@ inline ShapeInferContext::Strings ShapeInferContext::GetAttrStrings(const char*
   char c = {};
   size_t out = {};
   // first call to get the bytes needed
+  // 1. A status == nullptr means that ReadOpAttr was successful. A status != nullptr means failure.
+  // 2. The ReadOpAttr function should normally be called twice: once to get the needed buffer size (returns a status != nullptr), and a second time to actually read the ints (returns status == null on success).
+  // 3. This code tries a subtle optimization in the first call to ReadOpAttr. It passes in a buffer (&i) of size 1 just in case there is only 1 int. In this case, status == nullptr and we need to return {i}.
   auto status = ort_api_->ReadOpAttr(attr, ORT_OP_ATTR_STRINGS, &c, sizeof(char), &out);
   if (status) {
     std::vector<char> chars(out, '\0');
@@ -2115,6 +2154,9 @@ inline ShapeInferContext::Strings ShapeInferContext::GetAttrStrings(const char*
     }
     return strings;
   } else {
+    if (out == 0u) {
+      return {};
+    }
     return {std::string{c}};
   }
 }

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/onnxruntime_float16.h

@@ -75,8 +75,6 @@ struct Float16Impl {
   static constexpr uint16_t kNegativeInfinityBits = 0xFC00U;
   static constexpr uint16_t kPositiveQNaNBits = 0x7E00U;
   static constexpr uint16_t kNegativeQNaNBits = 0xFE00U;
-  static constexpr uint16_t kEpsilonBits = 0x4170U;
-  static constexpr uint16_t kMinValueBits = 0xFBFFU;  // Minimum normal number
   static constexpr uint16_t kMaxValueBits = 0x7BFFU;  // Largest normal number
   static constexpr uint16_t kOneBits = 0x3C00U;
   static constexpr uint16_t kMinusOneBits = 0xBC00U;
@@ -364,9 +362,6 @@ struct BFloat16Impl {
   static constexpr uint16_t kNegativeInfinityBits = 0xFF80U;
   static constexpr uint16_t kPositiveQNaNBits = 0x7FC1U;
   static constexpr uint16_t kNegativeQNaNBits = 0xFFC1U;
-  static constexpr uint16_t kSignaling_NaNBits = 0x7F80U;
-  static constexpr uint16_t kEpsilonBits = 0x0080U;
-  static constexpr uint16_t kMinValueBits = 0xFF7FU;
   static constexpr uint16_t kMaxValueBits = 0x7F7FU;
   static constexpr uint16_t kRoundToNearest = 0x7FFFU;
   static constexpr uint16_t kOneBits = 0x3F80U;

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Headers/onnxruntime_session_options_config_keys.h

@@ -269,6 +269,9 @@ static const char* const kOrtSessionOptionEpContextEmbedMode = "ep.context_embed
 // in case user need to merge/connect multiple EPContext nodes in one model
 static const char* const kOrtSessionOptionEpContextNodeNamePrefix = "ep.context_node_name_prefix";
 
+// Share EP related resources across EPs
+static const char* const kOrtSessionOptionShareEpContexts = "ep.share_ep_contexts";
+
 // Gemm fastmath mode provides fp32 gemm acceleration with bfloat16 based matmul.
 // Option values:
 // - "0": Gemm FastMath mode is not enabled. [DEFAULT]
@@ -279,3 +282,10 @@ static const char* const kOrtSessionOptionsMlasGemmFastMathArm64Bfloat16 = "mlas
 // Refer to MatMulNBits op schema for more details.
 // If not provided, default is 4.
 static const char* const kOrtSessionOptionsQDQMatMulNBitsAccuracyLevel = "session.qdq_matmulnbits_accuracy_level";
+
+// THIS OPTION IS NOT A REGULAR SESSION OPTION SINCE IT CAN BE MODIFIED AT ANY TIME
+// Meant to be used with SetEpDynamicOptions
+// Specify the type of workload for this session.
+// “Default”: OS determines the scheduling priority and processor performance to service this workload. [Default]
+// “Efficient”: OS treats this workload is efficiency oriented with low scheduling priority and efficient processor performance.
+static const char* const kOrtEpDynamicOptionsWorkloadType = "ep.dynamic.workload_type";

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64/onnxruntime.framework/Info.plist

@@ -9,9 +9,9 @@
 	<key>CFBundleIdentifier</key>
 	<string>com.microsoft.onnxruntime</string>
 	<key>CFBundleVersion</key>
-	<string>1.19.2</string>
+	<string>1.20.0</string>
 	<key>CFBundleShortVersionString</key>
-	<string>1.19.2</string>
+	<string>1.20.0</string>
 	<key>CFBundleSignature</key>
 	<string>????</string>
 	<key>CFBundlePackageType</key>

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64_x86_64-simulator/onnxruntime.framework/Headers/coreml_provider_factory.h

@@ -31,9 +31,14 @@ enum COREMLFlags {
   // Create an MLProgram. By default it will create a NeuralNetwork model. Requires Core ML 5 or later.
   COREML_FLAG_CREATE_MLPROGRAM = 0x010,
 
+  // Exclude ANE as sometimes this decrease performance
+  // https://developer.apple.com/documentation/coreml/mlcomputeunits?language=objc
+  // there are four compute units:
+  // MLComputeUnitsCPUAndNeuralEngine|MLComputeUnitsCPUAndGPU|MLComputeUnitsCPUOnly|MLComputeUnitsAll
+  COREML_FLAG_USE_CPU_AND_GPU = 0x020,
   // Keep COREML_FLAG_LAST at the end of the enum definition
   // And assign the last COREMLFlag to it
-  COREML_FLAG_LAST = COREML_FLAG_CREATE_MLPROGRAM,
+  COREML_FLAG_LAST = COREML_FLAG_USE_CPU_AND_GPU,
 };
 
 #ifdef __cplusplus

package/Plugins/iOS~/onnxruntime.xcframework/ios-arm64_x86_64-simulator/onnxruntime.framework/Headers/onnxruntime_c_api.h

@@ -38,7 +38,7 @@
  *
  * This value is used by some API functions to behave as this version of the header expects.
  */
-#define ORT_API_VERSION 19
+#define ORT_API_VERSION 20
 
 #ifdef __cplusplus
 extern "C" {
@@ -304,6 +304,7 @@ ORT_RUNTIME_CLASS(Op);
 ORT_RUNTIME_CLASS(OpAttr);
 ORT_RUNTIME_CLASS(Logger);
 ORT_RUNTIME_CLASS(ShapeInferContext);
+ORT_RUNTIME_CLASS(LoraAdapter);
 
 #ifdef _WIN32
 typedef _Return_type_success_(return == 0) OrtStatus* OrtStatusPtr;
@@ -621,6 +622,7 @@ typedef struct OrtMIGraphXProviderOptions {
   const char* migraphx_save_model_path;              // migraphx model path name
   int migraphx_load_compiled_model;                  // migraphx int8 cal table. Default 0 = false, noznero = true
   const char* migraphx_load_model_path;              // migraphx model path name
+  bool migraphx_exhaustive_tune;                     // migraphx tuned compile  Default = false
 } OrtMIGraphXProviderOptions;
 
 /** \brief OpenVINO Provider Options
@@ -643,7 +645,7 @@ typedef struct OrtOpenVINOProviderOptions {
    * Valid settings are one of: "CPU_FP32", "CPU_FP16", "GPU_FP32", "GPU_FP16"
    */
   const char* device_type;
-  unsigned char enable_npu_fast_compile;  ///< 0 = disabled, nonzero = enabled
+  unsigned char enable_npu_fast_compile;
   const char* device_id;
   size_t num_of_threads;  ///< 0 = Use default number of threads
   const char* cache_dir;  // path is set to empty by default
@@ -3649,10 +3651,17 @@ struct OrtApi {
    *     - "73"
    *     - "75"
    *   "device_id": The ID of the device to use when setting 'htp_arch'. Defaults to "0" (for single device).
-       "enable_htp_fp16_precision": Only used for float32 model.
-       Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision.
-         - "0": Default. With fp32 precision.
-         - "1": With fp16 precision.
+   *   "enable_htp_fp16_precision": Used for float32 model for HTP backend.
+   *   Enable the float32 model to be inferenced with fp16 precision. Otherwise, it will be fp32 precision.
+   *     - "0": With fp32 precision.
+   *     - "1": Default. With fp16 precision.
+   *   "enable_htp_weight_sharing": Enable QNN weight sharing feature while compiling multiple graphs into one QNN context.
+   *     - "0": Default. Disabled.
+   *     - "1": Enabled.
+   *   "offload_graph_io_quantization": Offload graph input quantization and graph output dequantization to another
+   *   execution provider (typically CPU EP).
+   *     - "0": Default. Disabled. QNN EP will handle quantization and dequantization of graph I/O.
+   *     - "1": Enabled.
    *
    * SNPE supported keys:
    *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",
@@ -3778,7 +3787,7 @@ struct OrtApi {
 
   /** \brief Release an OrtCANNProviderOptions
    *
-   * \param[in] the pointer of OrtCANNProviderOptions which will been deleted
+   * \param[in] input The pointer of OrtCANNProviderOptions which will been deleted
    *
    * \since Version 1.13.
    */
@@ -4666,6 +4675,82 @@ struct OrtApi {
                   _In_reads_(num_external_initializer_files) char* const* external_initializer_file_buffer_array,
                   _In_reads_(num_external_initializer_files) const size_t* external_initializer_file_lengths,
                   size_t num_external_initializer_files);
+
+  /** \brief Create an OrtLoraAdapter
+   *
+   * The function attempts to locate file specified by adapter_file_path, read it and create an OrtLoraAdapter
+   * instance. The adapter_file_path should be a valid path to a file that contains a valid Lora Adapter
+   * format. The function attempts to validate the format at load time. The file will always be memory mapped, unless
+   * the platform does not support memory mapping, in which case the file will be read into memory.
+   *
+   * \param[in] adapter_file_path adapter file path.
+   * \param[in] allocator optional pointer to a device allocator. If specified
+   *            data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU.
+   *            The data would still be copied to device if required by the model at inference time.
+   * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with
+   *                  OrtApi::ReleaseLoraAdapter.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(CreateLoraAdapter, const ORTCHAR_T* adapter_file_path, _In_ OrtAllocator* allocator,
+                  _Outptr_ OrtLoraAdapter** out);
+
+  /** \brief Create an OrtLoraAdapter
+   *
+   * The function copies the bytes from the array and creates an OrtLoraAdapter instance.
+   *
+   *
+   * \param[in] bytes pointer to a valid Lora Adapter format buffer.
+   * \param[in] num_bytes length of bytes buffer.
+   * \param[in] allocator optional pointer to a device allocator. If specified
+   *            data is copied to the device at some point before Run() is invoked. If nullptr, data stays on CPU.
+   *            The data would still be copied to device if required by the model at inference time.
+   * \param[out] out A pointer to a newly created OrtLoraAdapter instance. Must be released with
+   *                  OrtApi::ReleaseLoraAdapter.
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(CreateLoraAdapterFromArray, _In_ const void* bytes, size_t num_bytes, _In_ OrtAllocator* allocator,
+                  _Outptr_ OrtLoraAdapter** out);
+
+  /** \brief Release an ::OrtLoraAdapter obtained from OrtApi::CreateLoraAdapter
+   */
+  ORT_CLASS_RELEASE(LoraAdapter);
+
+  /** \brief Add the Lora Adapter to the list of active adapters.
+   *
+   * The function adds the Lora Adapter to the list of active adapters. The Lora Adapter must be created with
+   * OrtApi::CreateLoraAdapter or FromArray. The Lora Adapter will be used by the session to run the model.
+   * The instance of the OrtRunOptions can then be used to customize the Run() calls.
+   * More than one OrtLoraAdapter can be active at the same time. Lora Parameters that belong to different
+   * Lora adapters that will be active at the same time must not overlap.
+   * This setting does not affect RunWithBinding.
+   *
+   * \param[in] options OrtRunOptions instance
+   * \param[in] adapter OrtLoraAdapter instance
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(RunOptionsAddActiveLoraAdapter, _Inout_ OrtRunOptions* options, _In_ const OrtLoraAdapter* adapter);
+
+  /// @}
+  /// \name OrtEpDynamicOptions
+  /// @{
+
+  /** \brief Set DynamicOptions for EPs (Execution Providers)
+   *
+   * Valid options can be found in `include\onnxruntime\core\session\onnxruntime_session_options_config_keys.h`
+   * Look for `kOrtEpDynamicOptions`
+   *
+   * \param[in] sess OrtSession
+   * \param[in] keys Array of null terminated UTF8 encoded strings of EP dynamic option keys
+   * \param[in] values Array of null terminated UTF8 encoded string of EP dynamic option values
+   * \param[in] kv_len Number of elements in the keys and values arrays
+   *
+   * \snippet{doc} snippets.dox OrtStatus Return Value
+   */
+  ORT_API2_STATUS(SetEpDynamicOptions, _Inout_ OrtSession* sess, _In_reads_(kv_len) const char* const* keys,
+                  _In_reads_(kv_len) const char* const* values, _In_ size_t kv_len);
 };
 
 /*