SWA bounded replay · encoder side

Rebuilding the SWA KV after a prefix hit

A synthetic example with 6 layers (V4.1-Flash has L = 40). One cell = one window of W = 128 tokens; not to scale.

vLLM flag: --[no-]swa-bounded-replay, on by default for DeepSeek-V4.1.

SWA KV recomputed (exact) SWA KV recomputed (bounded replay) global KV, cached skipped SWA KV not cached