feat(pflash): prefill compress up to 128k -> 2-12× prefill (content-dependent), decode at parity#274

Open

dusterbloom wants to merge 16 commits into

Luce-Org:mainfrom

dusterbloom:feat/pflash-drafter-ee7

Commits on May 28, 2026

feat(pflash): ee7 early-exit drafter + anchor-transitive cascade + bug-42 tail-capture guard
dusterbloom
committed
refactor(pflash): rename DFLASH_COMPRESS_* → PFLASH_COMPRESS_* (cascade env vars)
dusterbloom
committed
fix(pflash): adaptive anchor_radius eliminates 64K NIAH cliff
dusterbloom
committed
bench: add eval_quality_compare.py for LongBench F1 regression detection
dusterbloom
committed
feat(qwen35): derive scalars from weights, assert vs GGUF metadata
dusterbloom
committed
feat(pflash): adaptive composition via per-request fa_window override
dusterbloom
committed
feat(pflash): PFLASH_*/DFLASH_* env-var dual aliasing + transitive cascade default-on
dusterbloom
committed
refactor(pflash): extract compress_cfg_from_env, kill qwen35/qwen3 path drift
dusterbloom
committed
chore(pflash): move narrative comments to docs/, trim mega-blocks
dusterbloom
committed
fix(server): append closed <think> prefill in Jinja renderer when thinking is off
dusterbloom
committed
fix(chat_template): gate closed-think prefill injection to Qwen3 arch only
dusterbloom
committed
refactor(c2-gate): wire c2_spec_decode_permitted into qwen35_backend
dusterbloom
committed

Commits on May 29, 2026

feat(pflash): effective-size admission gate + keep-ratio guard (keep default 0.10)
dusterbloom
committed

Commits on May 31, 2026