Optimize GLM-5.3 B12X and DFlash2 MXFP8 serving - #491

Closed

voipmonitor wants to merge 13 commits into

local-inference-lab:dev/jovian-judgementfrom

voipmonitor:codex/glm53-dflash2-mxfp8-20260828

Commits on Aug 28, 2026

Bind B12X KDA from caller scratch per invocation
voipmonitor
committed
3328207
Route GLM C4 prefill through DeepGEMM
voipmonitor
committed
534dc16
Parallelize GLM C4 prefill pool updates
voipmonitor
committed
18dd85c
Dispatch GLM mHC by token batch size
voipmonitor
committed
159ca47
perf(glm5next): dispatch KDA decode by request mode
voipmonitor
committed
28ee5e2
fix(glm5next): keep KDA backend stable per server
voipmonitor
committed
a4b04ee
Merge GLM MoE router gate correction
voipmonitor
committed
e4121e8
fix(cudagraph): retain custom-op capture resources
voipmonitor
committed
479f2ee
test(glm5next): match runtime QSA page contract
voipmonitor
committed
0fd7184
feat(dflash2): execute serialized MXFP8 projections natively
voipmonitor
committed
277fb5e
fix(attention): preserve draft-specific KV cache formats
voipmonitor
committed
1f9a229
fix(glm5next): pin multimodal processor revision
voipmonitor
committed
608010a
fix: validate DFlash graph resource contracts
voipmonitor
committed
b77333c