[AgentX] Mirror GB300 MTP tuning (#2571) to GB200 with 4-GPU->8-GPU workers - #2619

Closed

xinli-sw wants to merge 11 commits into

mainfrom

agentx/gb200-mtp2-from-gb300-2571

Commits on Aug 16, 2026

[AgentX] Mirror GB300 MTP tuning (#2571) to GB200 with 4-GPU->8-GPU workers

xinli-sw
and
claude
committed
668e916
Fill in pr-link for GB200 mtp2 changelog entry

xinli-sw
and
claude
committed
445d98f
Switch GB200 mtp2 recipes to stock nightly image + regular MegaMoE

xinli-sw
and
claude
committed
0f144af
Revert "Switch GB200 mtp2 recipes to stock nightly image + regular MegaMoE"
xinli-sw
committed
ecbb141
GB200 mtp2 disagg: lower Mooncake segment to 160GB, halve conc points

xinli-sw
and
claude
committed
97644bd
Fix decode CUDA-graph capture size on GB200 c128 disagg recipe

xinli-sw
and
claude
committed
0cf99d4
Lower GB200 c128 prefill gpu-memory-utilization to leave OOM headroom

xinli-sw
and
claude
committed
3914b27
GB200 mtp2 disagg: 140GB Mooncake segment, add 1p1d c256 point

xinli-sw
and
claude
committed
c082cd5

Commits on Aug 17, 2026

Merge branch 'main' into agentx/gb200-mtp2-from-gb300-2571
cquil11
authored
b7c1a1c
Update perf-changelog.yaml
cquil11
authored
907a4f2

Commits on Aug 19, 2026

Merge branch 'main' into agentx/gb200-mtp2-from-gb300-2571
Oseltamivir
authored
e6ea2bd