Module paged_executor 

Model executor that uses PagedAttention KV cache.

Unlike MockModelExecutor (which ignores KV cache), this executor:

• Writes K/V vectors to paged blocks during prefill and decode
• Reads K/V through block table indirection for attention
• Produces logits via the paged attention output

Uses identity projections (Q=K=V=input embedding) for deterministic, verifiable behavior without model weights.

Structs

PagedAttentionExecutor

A model executor that actually uses paged KV cache for attention.

PagedExecutorConfig

Configuration for the paged attention executor.