Optimize GLM-5.2-MXFP4 SGLang Agentic Performance on MI355X - #2900

Open

jiejingzhangamd wants to merge 18 commits into

mainfrom

dev/jiejing/glm5.2-sglang-opt-0908

Commits on Sep 8, 2026

update config for glm5.2-mxfp4 on mi355x gpu
lxgsbqylbk
committed
3050458
update pr link
lxgsbqylbk
committed
5361b95
Merge remote-tracking branch 'origin/main' into dev/limou/glm5.2-sglang-opt-0908
jiejingzhangamd
committed
09c2e8b
Update changelog link for replacement PR

jiejingzhangamd
and
cursoragent
committed
c8c7ca7
Update changelog link for upstream PR

jiejingzhangamd
and
cursoragent
committed
4937165
Merge remote-tracking branch 'origin/main' into dev/jiejing/glm5.2-sglang-opt-0908
jiejingzhangamd
committed
52835e9

Commits on Sep 9, 2026

Merge branch 'main' into dev/jiejing/glm5.2-sglang-opt-0908
chunfangamd
committed
3ac32bd

Commits on Sep 10, 2026

Commits on Sep 11, 2026

Merge branch 'main' into dev/jiejing/glm5.2-sglang-opt-0908
jiejingzhangamd
authored
f2e9fff
Bump the GLM-5.2 MI355X SGLang daily image to 20260910.

jiejingzhangamd
and
cursoragent
committed
10ffc71
Lower the TP HiCache host-pool ratio from 1.5 to 1.0.

jiejingzhangamd
and
cursoragent
committed
a47d5ae
Keep HiCache write_through_selective instead of restoring write_through.

jiejingzhangamd
and
cursoragent
committed
8e7a132
Rename --cuda-graph-max-bs for the v0.5.19 image.

jiejingzhangamd
and
cursoragent
committed
81d4111
Restore HiCache write-through policy

jiejingzhangamd
and
cursoragent
committed
aed71d7
Merge remote-tracking branch 'origin/main' into dev/jiejing/glm5.2-sglang-opt-0908
jiejingzhangamd
committed
f0d8627
Use canonical decode CUDA graph limit

jiejingzhangamd
and
cursoragent
committed
b43627e
Switch GLM-5.2 DSA prefill/decode to Triton.

jiejingzhangamd
and
cursoragent
committed
45c7ebf