[NPU]Move part DynamicGraph func toDynamicPipeline and refactor dynamic graph

[DanLiu2Intel]

Details:

Refactor DynamicGraph interface, move vm_execution logic and add static predict_shape func to DynamicPipeline
Details:

• Add DynamicPipeline::execute_vm_runtime to replace dynamicGraph->execute in DynamicPipeline::push().
• Add DynamicPipeline::predict_output_shape to replace dynamicGraph->predict_output_shape in ZeroDynamicInferRequest::infer_predict_shapes. This logic is not moved to DynamicPipeline::push() in this PR to prevent DynamicPipeline from including _userOutputTensors and _userinputTensors from ZeroDynamicInferRequest.
• Consolidate IDynamicGraph::GraphArgument with DynamicGraph::GraphArgumentsImpl to intel_npu::GraphArgument in dynamic_pipeline. Move DynamicGraph::MemRefTypeImp to the intel_npu::MemRefTypeImp in dynamic_pipeline. Move IDynamicGraph::MemRefType to intel_npu::MemRefType in utils/vm/dynamic_arguments file.
• Remove the binding methods from GraphArguments and DynamicGraphImpl, and merge the remaining functionality of DynamicGraphImpl into GraphArguments.
• Update igraph::get_handle to return pointers to npu_vm_runtime_handle_t and ze_graph_handle_t. Add num_subgraphs metadata to deprecate and remove IDynamicGraph. Update DynamicGraph to inherit from IGraph.
• Include openvino_npu_vm_utils into openvino_npu_common to resolve the ov_npu_unit_tests linking issue.
• Remove npu_vm_runtime_api target and move npu_vm_runtime.hpp to vm/utils

Tickets:

• CVS-183328

AI Assistance:

• AI assistance used: no / yes
• If yes, summarize how AI was used and what human validation was performed (build/tests/manual checks).

[Copilot]

Pull request overview

This PR refactors the Intel NPU plugin “dynamic graph” path by removing the IDynamicGraph interface, introducing a shared GraphArguments/MemRefType API in common, and moving VM-runtime execution + shape prediction logic into the Level Zero dynamic pipeline. Build wiring is updated so VM utils are a standalone static library target (openvino::npu_vm_utils) and linked where needed.

Changes:

• Replace IDynamicGraph with IGraph extensions (get_vm_runtime_handle(), get_num_subgraphs()), and simplify DynamicGraph accordingly.
• Introduce dynamic_graph_arguments.hpp (+ impl header) and migrate argument/VM MemRef tracking to the backend pipeline.
• Update CMake targets to build/link/install openvino_npu_vm_utils as a static library and link it from dependent components.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
src/plugins/intel_npu/src/utils/src/vm/CMakeLists.txt	Switch VM utils from OBJECT → STATIC lib, add alias and install rule
src/plugins/intel_npu/src/compiler_adapter/src/dynamic_graph.cpp	Remove per-call binding/execute/predict logic; expose VM runtime handle
src/plugins/intel_npu/src/compiler_adapter/include/dynamic_graph.hpp	Move from IDynamicGraph to IGraph, drop old dynamic-graph-specific APIs
src/plugins/intel_npu/src/compiler_adapter/CMakeLists.txt	Stop consuming VM utils as object files; link openvino::npu_vm_utils
src/plugins/intel_npu/src/common/src/igraph.cpp	Add default get_vm_runtime_handle() / get_num_subgraphs() implementations
src/plugins/intel_npu/src/common/include/intel_npu/common/igraph.hpp	Extend IGraph interface with VM-runtime handle + subgraph count
src/plugins/intel_npu/src/common/src/dynamic_graph_arguments.cpp	Move MemRef/GraphArguments implementation out of removed IDynamicGraph
src/plugins/intel_npu/src/common/include/intel_npu/common/dynamic_graph_arguments.hpp	New shared MemRefType + GraphArguments API
src/plugins/intel_npu/src/common/include/intel_npu/common/dynamic_graph_arguments_impl.hpp	New header-only impl for VM MemRef + execute params lifetime tracking
src/plugins/intel_npu/src/common/CMakeLists.txt	Link common against VM utils
src/plugins/intel_npu/src/backend/src/zero_dynamic_pipeline.cpp	Implement VM-runtime execute + predict-shape in pipeline
src/plugins/intel_npu/src/backend/include/zero_dynamic_pipeline.hpp	Rework binding to use new GraphArguments; add predict-shape API
src/plugins/intel_npu/src/backend/src/zero_dynamic_infer_request.cpp	Route predict-shape through pipeline static helper; use new MemRefType
src/plugins/intel_npu/src/backend/include/zero_dynamic_infer_request.hpp	Update signatures to new MemRefType
src/plugins/intel_npu/src/backend/src/zero_device.cpp	Detect dynamic graphs via get_vm_runtime_handle() rather than RTTI
src/plugins/intel_npu/src/common/include/intel_npu/common/idynamic_graph.hpp	Remove obsolete interface

Comments suppressed due to low confidence (2)

src/plugins/intel_npu/src/common/src/dynamic_graph_arguments.cpp:138

• [BLOCKER] GraphArguments::setArgumentProperties() can read past the end of the provided strides vector.

In assign_slot(), when _dimsCount==0 the code resizes slot._strides to strides.size(), but then unconditionally loops i < slot._dimsCount (== sizes.size()) and indexes strides[i]. If strides.size() != sizes.size() (e.g. empty strides for scalar/unknown-strides tensors), this is out-of-bounds and will crash in the dynamic pipeline/predict-shape paths.

Please validate strides.size() == sizes.size() (or treat empty strides as contiguous and compute them from sizes), and always size slot._strides to sizes.size().
src/plugins/intel_npu/src/utils/src/vm/CMakeLists.txt:27

• [MEDIUM] openvino_npu_vm_utils was converted from an OBJECT library to a standalone STATIC library, but it still doesn't declare its link dependencies.

npu_vm_runtime_api.cpp uses ov::util (make_plugin_library_name/load_shared_object/get_symbol). With a STATIC library, relying on every consumer to link the right OpenVINO utility targets is brittle and can cause link-order failures when openvino::npu_vm_utils is consumed independently.

Please add an explicit target_link_libraries(openvino_npu_vm_utils ...) (at least openvino::util; and any other required libs) to make the target self-contained.

add_library(${TARGET_NAME} STATIC ${SOURCES})
add_library(openvino::npu_vm_utils ALIAS ${TARGET_NAME})

set_target_properties(${TARGET_NAME} PROPERTIES
    INTERPROCEDURAL_OPTIMIZATION_RELEASE ${ENABLE_LTO}
    EXPORT_NAME npu_vm_utils)

ov_add_clang_format_target(${TARGET_NAME}_clang FOR_TARGETS ${TARGET_NAME})

target_include_directories(${TARGET_NAME}
    PUBLIC
        $<BUILD_INTERFACE:${NPU_UTILS_SOURCE_DIR}/include>
        $<TARGET_PROPERTY:openvino::util,INTERFACE_INCLUDE_DIRECTORIES>
        $<TARGET_PROPERTY:openvino::core::dev,INTERFACE_INCLUDE_DIRECTORIES>
        $<TARGET_PROPERTY:openvino::runtime::dev,INTERFACE_INCLUDE_DIRECTORIES>
        $<TARGET_PROPERTY:npu_vm_runtime_api,INTERFACE_INCLUDE_DIRECTORIES>
        $<TARGET_PROPERTY:openvino_npu_zero_utils,INTERFACE_INCLUDE_DIRECTORIES>
)

[Copilot]

Pull request overview

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (2)

src/plugins/intel_npu/src/common/src/dynamic_graph_arguments.cpp:138

• [BLOCKER] GraphArguments::setArgumentProperties() can read past the end of strides (and write past _strides) when the caller passes an empty strides vector (common for continuous tensors / tensors with get_strides().empty()). slot._strides is resized to strides.size() (possibly 0), but the subsequent loop indexes strides[i] for i < _dimsCount.

Fix: require strides.empty() || strides.size() == sizes.size(), and when strides.empty() compute contiguous element-strides (e.g., fill _sizes then call slot.updateStride()), otherwise resize _strides to sizes.size() before filling.
src/plugins/intel_npu/src/common/src/dynamic_graph_arguments.cpp:70

• [HIGH] MemRefType::set() assumes tensor->get_strides() has rank elements and indexes it unconditionally. For OpenVINO tensors, get_strides() can be empty (e.g., sub-byte element types), which would cause out-of-bounds access here.

Fix: if tensor->get_strides().empty(), populate _strides via updateStride() (contiguous row-major) instead of indexing the empty stride vector; otherwise validate strides.size() == _dimsCount before reading.

[Copilot]

Pull request overview

Copilot reviewed 14 out of 14 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (1)

src/plugins/intel_npu/src/compiler_adapter/src/dynamic_graph.cpp:131

• [BLOCKER] prepare_metadata() passes upperBound.data() from a vector that is only reserved (size==0). upperBound.data() will be nullptr, so npuVMRuntimeGetMetadata may write through a null pointer / out of bounds.

Allocate the buffer with the required size (e.g., ZE_MAX_GRAPH_ARGUMENT_DIMENSIONS_SIZE) before passing its data pointer.

        ze_graph_argument_properties_3_t arg;
        ze_graph_argument_metadata_t meta;
        std::vector<int64_t> upperBound;
        upperBound.reserve(ZE_MAX_GRAPH_ARGUMENT_DIMENSIONS_SIZE);
        if (npuVMRuntimeGetMetadata(_engine, i, &arg, &meta, upperBound.data()) != NPU_VM_RUNTIME_RESULT_SUCCESS) {

[Copilot]

Pull request overview

Copilot reviewed 28 out of 29 changed files in this pull request and generated 2 comments.