perf(qwen35): use in-place GDN state write in pure AR#473

Open

dusterbloom wants to merge 9 commits into

Luce-Org:mainfrom

dusterbloom:perf/qwen35-gdn-inplace-consumer

Commits on Jul 9, 2026

perf(qwen35): use in-place GDN state write in pure AR
dusterbloom
committed
perf(decode): fuse silu+mul -> swiglu_split in DeltaNet gated output
dusterbloom
committed
perf(qwen35): gate FWHT rotation for q4_0 KV + skip repeat_back in pure AR
dusterbloom
committed
perf(qwen35): cont concat inputs for fast concat_cont path
dusterbloom
committed
perf(qwen35): graph reuse in AR decode — skip rebuild within 256-token FA bucket
dusterbloom
committed
perf(qwen35): reduce concat cont overhead — materialize qkv_T once
dusterbloom
committed
perf(qwen35): recover dense AR replay fast path
dusterbloom
committed
fix(qwen35): scope AR props-skip to target compute; gate reuse to dense; K-basis in disk-cache salt; CPU GDN inplace guard
dusterbloom
committed

Commits on Jul 10, 2026

fix(qwen35): address AR reuse and CPU GDN reviews
dusterbloom
committed