Skip to content

npu: the dense-tail cache is on by default from a 128 MiB tail — Qwen3.5-2B decode 485–520 → 362–367 ms per token, same text (Refs #1785) - #2075

Merged
bong-water-water-bong merged 5 commits into
mainfrom
cl-flm/1785-ffn-cache-floor
Nov 11, 2026
Merged

bong-water-water-bong merged 5 commits into
mainfrom
cl-flm/1785-ffn-cache-floor

Commits

  1. Commits on Nov 11, 2026