[NPUW] Moving passes from llm_compiled_model to separate files

[AlexanderKalistratov]

Extract all inline OpenVINO passes and transformation logic from llm_compiled_model.cpp and llm_compiled_model_utils.cpp into dedicated source files. This PR intentionally doesn't change names or namespaces and not doing any other changes. Just moving code.

All common transformation are put into npuw_transformations/ folder:

• ConvertKVCacheToPrecision: convert_kvcache_to_precision.cpp/.hpp
• DecomposeGQA: decompose_gqa.cpp/.hpp
• Phi3SlidingMask: phi3_sliding_mask.cpp/.hpp
• PatchPhi3SlidingMask: patch_phi3_sliding_mask.cpp/.hpp
• ReshapeToStatic: reshape_to_static.cpp/.hpp
• ReshapeSlicedHeadToStatic: reshape_sliced_head_to_static.cpp/.hpp
• SliceOutEmbeds: slice_out_embeds.cpp/.hpp
• LoraStatefulToStatelessPass: lora_stateful_to_stateless.cpp/.hpp
• OptimizeValueTensors: optimize_value_tensors.cpp/.hpp (previously in llm_compiled_model_utils.cpp)
• kv_axes_position.hpp — KVAxesPosition struct (previously in anonymous namespace in llm_compiled_model.hpp)

All code related to embedding models are moved to embedding/ folder:

• RemoveEmptyKVInputs: remove_empty_kv_inputs.cpp/.hpp
• RedirectNewKvToOutput: redirect_new_kv_to_output.cpp/.hpp
• embedding_infer_request.cpp/.hpp and prepare_embedding_model.cpp/.hpp are moved

MoE related transformation are moved to moe_transformations/ folder:

• ApplyMoEDeviceRoutedTransforms: apply_moe_device_routed_transforms.cpp/.hpp
• The rest of MoE related code was already there

Whisper related transformations are mode to whisper/ folder:

• PrepareWhisperPrefillModel, PrepareWhisperKVCacheModel: prepare_whisper_model.cpp/.hpp (prevoiusly were in llm_compiled_model_utils.cpp)
• whisper_infer_request.cpp/.hpp were moved

AI Assistance:

AI assistance used yes
AI moved most of the code, reviewed and check verified manually

[Copilot]

Pull request overview

This PR refactors the Intel NPUW plugin implementation by extracting previously inline LLM/Whisper/Embedding/MoE transformation passes and related logic from the large llm_compiled_model*.cpp/hpp and llm_compiled_model_utils.* files into dedicated translation units under clearer subfolders (npuw_transformations/, embedding/, whisper/, moe_transformations/). This primarily improves code organization and compilation-unit boundaries while aiming to preserve existing behavior.

Changes:

• Moved multiple OpenVINO ModelPass / matcher-pass implementations into new dedicated .cpp/.hpp files under npuw_transformations/, embedding/, whisper/, and moe_transformations/.
• Updated LLMCompiledModel and unit tests to include the new headers and compile the new sources.
• Introduced KVAxesPosition as a shared struct header (kv_axes_position.hpp) instead of an anonymous-namespace definition.

Reviewed changes

Copilot reviewed 39 out of 40 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/plugins/intel_npu/tests/unit/npuw/transpose_vt.cpp	Adjusts includes to use the extracted OptimizeValueTensors header.
src/plugins/intel_npu/tests/unit/npuw/llm_compiled_model_factory_options_test.cpp	Updates includes to use extracted Whisper preparation passes.
src/plugins/intel_npu/tests/unit/CMakeLists.txt	Adds newly extracted plugin sources to the unit test target build.
src/plugins/intel_npu/src/plugin/npuw/whisper/whisper_infer_request.hpp	Fixes include paths after moving Whisper code into a subfolder.
src/plugins/intel_npu/src/plugin/npuw/whisper/whisper_infer_request.cpp	Fixes include paths after moving Whisper code into a subfolder.
src/plugins/intel_npu/src/plugin/npuw/whisper/prepare_whisper_model.hpp	New header for Whisper model preparation passes.
src/plugins/intel_npu/src/plugin/npuw/whisper/prepare_whisper_model.cpp	New implementation of Whisper model preparation transformations.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/slice_out_embeds.hpp	New header for SliceOutEmbeds pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/slice_out_embeds.cpp	New implementation for SliceOutEmbeds pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/reshape_to_static.hpp	New header for ReshapeToStatic pass and KVAxesPosition usage.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/reshape_to_static.cpp	New implementation for ReshapeToStatic logic.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/reshape_sliced_head_to_static.hpp	New header for ReshapeSlicedHeadToStatic pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/reshape_sliced_head_to_static.cpp	New implementation for reshaping sliced LM head inputs.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/phi3_sliding_mask.hpp	New header for Phi3SlidingMask transformation pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/phi3_sliding_mask.cpp	New implementation of Phi-3 sliding window attention mask patching.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/patch_phi3_sliding_mask.hpp	New wrapper pass header to apply Phi-3 sliding mask transformations.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/patch_phi3_sliding_mask.cpp	New wrapper pass implementation.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/optimize_value_tensors.hpp	New header for OptimizeValueTensors pass (SDPA decomposition + V tensor transpose).
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/optimize_value_tensors.cpp	New implementation for OptimizeValueTensors.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/lora_stateful_to_stateless.hpp	New header for LoRA stateful→stateless conversion pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/lora_stateful_to_stateless.cpp	New implementation for LoRA state conversion.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/kv_axes_position.hpp	New shared KVAxesPosition struct header.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/decompose_gqa.hpp	New header for DecomposeGQA pass.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/decompose_gqa.cpp	New implementation for GroupQueryAttention decomposition.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/convert_kvcache_to_precision.hpp	New header for KV-cache low-precision conversion and type selection helper.
src/plugins/intel_npu/src/plugin/npuw/npuw_transformations/convert_kvcache_to_precision.cpp	New implementation for KV-cache precision conversion + FP8 optimization helper.
src/plugins/intel_npu/src/plugin/npuw/moe_transformations/apply_moe_device_routed_transforms.hpp	New header for applying DEVICE_ROUTED MoE transformations.
src/plugins/intel_npu/src/plugin/npuw/moe_transformations/apply_moe_device_routed_transforms.cpp	New implementation applying MoE transforms in sequence.
src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model_utils.hpp	Removes pass class declarations that were moved out.
src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model_utils.cpp	Removes moved transformation logic, leaving shared helpers like has_input.
src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.hpp	Replaces anonymous KVAxesPosition with shared header include.
src/plugins/intel_npu/src/plugin/npuw/llm_compiled_model.cpp	Switches to includes of the extracted passes/transforms and removes inlined implementations.
src/plugins/intel_npu/src/plugin/npuw/embedding/remove_empty_kv_inputs.hpp	New header for embedding-model pass to drop empty KV inputs.
src/plugins/intel_npu/src/plugin/npuw/embedding/remove_empty_kv_inputs.cpp	New implementation for removing empty KV inputs in embedding models.
src/plugins/intel_npu/src/plugin/npuw/embedding/redirect_new_kv_to_output.hpp	New header for embedding-model pass to redirect new KV tensors to outputs.
src/plugins/intel_npu/src/plugin/npuw/embedding/redirect_new_kv_to_output.cpp	New implementation for redirecting new KV tensors to outputs.
src/plugins/intel_npu/src/plugin/npuw/embedding/prepare_embedding_model.hpp	New header for embedding model preparation pass.
src/plugins/intel_npu/src/plugin/npuw/embedding/prepare_embedding_model.cpp	Updated implementation moved from old embedding utils, with updated include paths.
src/plugins/intel_npu/src/plugin/npuw/embedding/embedding_infer_request.hpp	Fixes include paths after moving embedding infer request into a subfolder.
src/plugins/intel_npu/src/plugin/npuw/embedding/embedding_infer_request.cpp	Fixes include paths after moving embedding infer request into a subfolder.

[Copilot]

[BLOCKER] This header uses uint32_t and std::size_t but does not include the headers that define them. It currently relies on transitive includes from OpenVINO headers, which is brittle and can break with include-order changes. Add explicit includes (e.g., <cstdint> and <cstddef> or <cstdint> and <cstdlib> as appropriate).

[Copilot]

[BLOCKER] This header uses uint32_t / std::size_t but doesn't include <cstdint> / <cstddef> (or equivalent) directly. Depending on transitive includes can make builds fail on different compilers or after unrelated include changes; please add the needed standard includes here.

[Copilot]

[MEDIUM] Several newly extracted transformation passes are now declared in the global namespace in public headers (starting with this one, and similarly in reshape_to_static.hpp, reshape_sliced_head_to_static.hpp, phi3_sliding_mask.hpp, convert_kvcache_to_precision.hpp, moe_transformations/apply_moe_device_routed_transforms.hpp, etc.). This increases collision risk and is inconsistent with existing NPUW transformation headers that scope passes under ov::npuw::pass / ov::npuw::... (e.g., src/plugins/intel_npu/src/plugin/npuw/moe_transformations/gather_to_2d_gather.hpp:10-42). Consider wrapping these class declarations in the appropriate ov::npuw (and/or ov::npuw::pass / ov::npuw::util) namespace to keep symbols scoped to the plugin and match established conventions.