[NPUW] Support KV cache i8 compression with pyramid attention

[esmirno]

Details:

• Add SDPACompressed pattern matcher to isolate i8 dequantized KV cache attention subgraphs (Convert -> Subtract -> Multiply -> Concat)
• Make Convert ops optional in SDPACompressed to handle models where past KV is already in the expected type
• Support both asymmetric (with zero-point Subtract) and symmetric (scale-only Multiply) dequantization
• Propagate DQ scale/zero-point parameters through pyramid attention model creation
• Add isDQScaleOrZPKey/isDQScaleOrZPValue utility functions for identifying DQ parameters by name
• Add unit tests for SDPACompressed pattern matching and pyramid attention with i8 KV cache

Tickets:

• EISW-218895

AI Assistance:

• AI assistance used: yes
• Copilot assisted with code implementation, pattern matcher design, and test creation.

[intelgaoxiong]

👍LGTM.
Anyway, the block-based KV cache will need to be extended in the future, but I hope to support fp16 KV cache as the first step.

[dmatveev]

A note for the future - instead of names, should we better tag it with rt_info?

[dmatveev]

what if it is empty? why we would have such a case?

[esmirno]

this case perfectly covers contract for this function API - while for entire flow it is right now not possible combination. I wouldnt call begin() for empty structure anyway, better to have more logs of course here or even controlled exceptions.

std::optional<PyramidAttention> PyramidAttention::from(const std::shared_ptr<ov::Model>& model) {
    // Validate and setup pyramid attention
    auto validation_result = validate_and_setup_pyramid_attention(model); <-- sequence dims gets parsed or died
    if (!validation_result) {
        return std::nullopt;
    }

[dmatveev]

would these inputs be handled by the pyramid's input ROI (view) logic the same way as past k/v len?

so no other extra handling required?

[esmirno]

I'm expecting so, quant params should match pyramid sequnce len for per-token quantization.

[CuriousPanCake]

If you only need to mark the matched nodes, consider obtaining them using Matcher::get_matched_values()

[esmirno]

@CuriousPanCake thanks for suggestion, for our need it looks that API cannot be used directly - get_matched_values() usages will require explicit filtering of any_input() nodes, that is useless overcomplication of simple code. So here we have verbose but direct control of what gets isolated versus most extensive filtering / deduplication

[CuriousPanCake]

What's the point of this class?

[intelgaoxiong]

I suppose the value of actual_context_length is the same with the original actual_context_length?