{"model":"deepseek-4.1-flash","title":"Sparse-MLA decode knob sweep - dead code found, +2-4% folded in","status":"stuck","summary":"VLLM_SPARSE_DECODE_MAXNREG is dead code (_decode_maxnreg_kwargs has no caller) - the earlier MAXNREG test measured nothing. Wired real knobs and swept on PP8/512k: HEAD_BLOCK_SIZE=32 gives consistent, bit-identical +2-4% (c1 53.0-\u003e54.3, 512k c8 45.7-\u003e47.4, c32 60.2-\u003e61.6) - folded in as PP8…","occurred_at":"2026-09-12T17:06:00Z","config":{"serving":"TP1xPP8 (Zanooda 13-patch stack), fp8_ds_mla","engine":"zanooda/vllm-sm80-ds41f:v41-sm80 + KV-layout intersection fix"},"benchmarks":[{"metric":"c1_decode_tok_s","value":54.3,"unit":"tok/s","context":"PP8 HEAD_BLOCK_SIZE=32, 512-tok gen (53.0 baseline)","note":"+2.4%, bit-identical output"},{"metric":"agg_decode_512k_tok_s","value":61.6,"unit":"tok/s","context":"PP8 c32 with HEAD_BLOCK_SIZE=32 (60.2 baseline)"}],"found":["Small knob gains confirm 512k decode is not KV-bandwidth-bound"],"fixed":["HEAD_BLOCK_SIZE=32 folded in as PP8 default; needle 4/4 at 128k/512k"],"tried":["VLLM_SPARSE_DECODE_MAXNREG (dead code - no effect)","TOPK_CHUNK_SIZE=16 (hurts c1: 49.1)"]}