[GLM-5.3 Flash] Serving optimizations: KDA/mHC fusion, DSA draft metadata graphs, prefill autotuning - #39422

Draft

JustinTong0323 wants to merge 10 commits into

mainfrom

xinyuan/glm53flash-sota-perf

Commits on Sep 14, 2026

Reduce register spills in dense KDA verification
JustinTong0323
committed
8fcf8ad
Use smaller state tiles for dense KDA verification
JustinTong0323
committed
93ef118
Parallelize small-token mHC post on GB300

JustinTong0323
and
mmangkad
committed
34f6a1f
Tune small-row gated RMSNorm on GB300
JustinTong0323
committed
1a6ac7b
Fuse small-token mHC across GLM layer boundaries

JustinTong0323
and
mmangkad
committed
d63e4ff
Autotune GLM FP8 experts on the first large target prefill
JustinTong0323
committed
32a5067
Capture pooled DSA draft metadata preparation

JustinTong0323
and
mmangkad
committed
992fe97
Enable qualified DSA draft metadata graphs by default

JustinTong0323
and
mmangkad
committed
b58a1ff
Extend qualified DSA draft metadata graphs to padded batches

JustinTong0323
and
mmangkad
committed
b05db76
[Perf] Reuse CPU token counts for ordinary KDA prefill
JustinTong0323
committed
36e3696