Chunk prefill on Key and Value.

[intelgaoxiong]

Details:

• Optimize TTFT by reducing redundant computing in SDPA on NPU.

Tickets:

• EISW-181644

[Copilot]

Pull Request Overview

This PR implements chunked Key-Value (KV) cache support for Large Language Model (LLM) inference by extending the existing chunked prefill functionality to include KV chunking alongside the existing query (Q) chunking.

• Introduces support for multiple prefill models with different past KV shapes for KV chunking
• Adds new configuration option NPUW_LLM_PREFILL_ENABLE_KV_CHUNK to control KV chunking behavior
• Implements KV cache management between chunks and tensor sharing optimizations

Reviewed Changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
llm_infer_request.hpp	Updates class structure to support multiple prefill requests and KV chunking state tracking
llm_infer_request.cpp	Implements core KV chunking logic including chunk-to-chunk KV cache updates and tensor management
llm_compiled_model.hpp	Extends model to support multiple prefill models for KV chunking scenarios
llm_compiled_model.cpp	Implements parallel compilation of multiple prefill models and KV chunking configuration
base_sync_infer_request.cpp	Updates tensor allocation tracking for remote tensors
npuw.cpp	Registers the new KV chunk configuration option
npuw_private_properties.hpp	Adds the new KV chunk enable property definition
npuw.hpp	Defines the configuration option macro for KV chunking

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}