added sdpa llm inference for qwen3.5 model support (#107)

* added modeling code from openvino-new-arch repo.

this is the code base dependendcy for supporing qwen3.5 model.

Signed-off-by: Zhang, Xiaolin <xiaolin.zhang@intel.com>

* added sdpa llm inference for qwen3.5 model support.

Signed-off-by: Zhang, Xiaolin <xiaolin.zhang@intel.com>

* added a unit test for LLMInferenceSDPAModule for Qwen3.5

Signed-off-by: Zhang, Xiaolin <xiaolin.zhang@intel.com>

* added compiler flag -DENABLE_OPENVINO_NEW_ARCH to control build behavior

to support both public openvino runtime and new arch for qwen3.5 support

Signed-off-by: Zhang, Xiaolin <xiaolin.zhang@intel.com>

---------

Signed-off-by: Zhang, Xiaolin <xiaolin.zhang@intel.com>

Author: xzhan34

cmake/features.cmake

@@ -11,6 +11,7 @@ option(ENABLE_TOOLS "Enable tools build" ON)
 option(ENABLE_GGUF "Enable support for GGUF format" ON)
 option(ENABLE_XGRAMMAR "Enable support for structured output generation with xgrammar backend" ON)
 option(ENABLE_DYNAMIC_WEIGHT_MANAGEMENT "Enable offloading model weights (load/release)" OFF)
+option(ENABLE_OPENVINO_NEW_ARCH "Enable OpenVINO new architecture for QWen3.5 etc models support" OFF)
 
 # Disable building samples for NPM package
 if(CPACK_GENERATOR STREQUAL "NPM")

samples/cpp/module_genai/CMakeLists.txt

@@ -46,4 +46,4 @@ if(POLICY CMP0135)
     cmake_policy(SET CMP0135 NEW)
 endif()
 
-add_subdirectory(comfyui)
+add_subdirectory(comfyui)

src/cpp/CMakeLists.txt

@@ -2,7 +2,8 @@
 # SPDX-License-Identifier: Apache-2.0
 #
 
-file(GLOB_RECURSE SOURCE_FILES CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/src/*.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/src/*.c")
+file(GLOB_RECURSE SOURCE_FILES "${CMAKE_CURRENT_SOURCE_DIR}/src/*.cpp" "${CMAKE_CURRENT_SOURCE_DIR}/src/*.c")
+
 list(APPEND SOURCE_FILES "${CMAKE_CURRENT_BINARY_DIR}/version.cpp")
 
 # Dependencies
@@ -131,6 +132,12 @@ set(TARGET_NAME_OBJ ${TARGET_NAME}_obj)
 
 add_library(${TARGET_NAME_OBJ} OBJECT ${SOURCE_FILES})
 
+if(ENABLE_OPENVINO_NEW_ARCH)
+    file(GLOB_RECURSE MODELING_SOURCES CONFIGURE_DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/src/modeling/*.cpp")
+    target_sources(${TARGET_NAME_OBJ} PRIVATE ${MODELING_SOURCES})
+    target_compile_definitions(${TARGET_NAME_OBJ} PRIVATE ENABLE_OPENVINO_NEW_ARCH)
+endif()
+
 target_include_directories(${TARGET_NAME_OBJ}
     PUBLIC "$<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/include>"
            "$<BUILD_INTERFACE:${CMAKE_CURRENT_BINARY_DIR}>"
@@ -154,6 +161,12 @@ target_include_directories(${TARGET_NAME_OBJ} SYSTEM PRIVATE "${safetensors.h_SO
 
 target_link_libraries(${TARGET_NAME_OBJ} PRIVATE openvino::runtime openvino::threading nlohmann_json::nlohmann_json minja ${YAML_CPP_TARGET} PRIVATE TBB::tbb)
 
+# modeling/ sources use internal OpenVINO dev-api headers (e.g. openvino/op/linear_attn.hpp)
+# openvino::core::dev exposes src/core/dev_api which contains these headers.
+if(TARGET openvino::core::dev)
+    target_link_libraries(${TARGET_NAME_OBJ} PRIVATE openvino::core::dev)
+endif()
+
 target_compile_features(${TARGET_NAME_OBJ} PRIVATE cxx_std_17)
 
 target_compile_definitions(${TARGET_NAME_OBJ} PRIVATE openvino_genai_EXPORTS)
@@ -190,7 +203,7 @@ endif()
 
 target_compile_features(${TARGET_NAME} INTERFACE cxx_std_17)
 
-# Add /bigobj flag for MSVC to handle large object files  
+# Add /bigobj flag for MSVC to handle large object files
 if(MSVC)
     target_compile_options(${TARGET_NAME} PRIVATE "/bigobj")
 endif()

src/cpp/src/gguf_utils/gguf_weight_finalizer.hpp

@@ -0,0 +1,39 @@
+// Copyright (C) 2023-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include <string>
+#include <unordered_map>
+
+#include <openvino/openvino.hpp>
+
+#include "gguf_utils/gguf.hpp"
+#include "modeling/ops/context.hpp"
+#include "modeling/weights/weight_finalizer.hpp"
+
+namespace ov {
+namespace genai {
+namespace gguf {
+
+class GGUFWeightFinalizer : public ov::genai::modeling::weights::WeightFinalizer {
+public:
+    GGUFWeightFinalizer(const std::unordered_map<std::string, ov::Tensor>& consts,
+                        const std::unordered_map<std::string, gguf_tensor_type>& qtypes);
+
+    ov::genai::modeling::weights::FinalizedWeight finalize(const std::string& name,
+                                                           ov::genai::modeling::weights::WeightSource& source,
+                                                           ov::genai::modeling::OpContext& ctx) override;
+
+private:
+    gguf_tensor_type resolve_qtype(const std::string& base_key) const;
+    std::string base_key_from_name(const std::string& name) const;
+
+    const std::unordered_map<std::string, ov::Tensor>& consts_;
+    const std::unordered_map<std::string, gguf_tensor_type>& qtypes_;
+    std::unordered_map<std::string, ov::Output<ov::Node>> cache_;
+};
+
+}  // namespace gguf
+}  // namespace genai
+}  // namespace ov

src/cpp/src/gguf_utils/gguf_weight_source.hpp

@@ -0,0 +1,34 @@
+// Copyright (C) 2023-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#pragma once
+
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include <openvino/openvino.hpp>
+
+#include "modeling/weights/weight_source.hpp"
+
+namespace ov {
+namespace genai {
+namespace gguf {
+
+class GGUFWeightSource : public ov::genai::modeling::weights::WeightSource {
+public:
+    explicit GGUFWeightSource(const std::unordered_map<std::string, ov::Tensor>& consts);
+
+    std::vector<std::string> keys() const override;
+    bool has(const std::string& name) const override;
+    const ov::Tensor& get_tensor(const std::string& name) const override;
+
+private:
+    const std::unordered_map<std::string, ov::Tensor>& consts_;
+    std::vector<std::string> keys_;  // Canonical names
+    std::unordered_map<std::string, std::string> canonical_to_gguf_;  // canonical -> gguf mapping
+};
+
+}  // namespace gguf
+}  // namespace genai
+}  // namespace ov