Fix Kimi K3 B200 AIPerf metrics configuration - #2569

Open

cquil11 wants to merge 38 commits into

mainfrom

agent/refresh-kimik3-b200-dspark-agentx

Commits on Aug 3, 2026

perf(agentx): add B200 Kimi K3 DSpark refresh
cquil11
committed
f94cabe
chore: link B200 DSpark refresh PR
cquil11
committed
256df24
feat: add b200 kimi k3 agentx profiles
adibarra
committed
9515848
fix: bypass unsupported b200 custom collectives
adibarra
committed
c5fa8da
fix: correct b200 dep launch accounting
adibarra
committed
88d3bb7
fix: force pynccl for cross-node tp
adibarra
committed
3cb4971
fix: disable cross-node latent tail fusion
adibarra
committed
94d677a
fix: disable cross-node allreduce fusion
adibarra
committed
349ddfd
fix: use viable b200 tp8pp2 profiles
adibarra
committed
48dffa1
fix: keep dspark on pp-free tp16
adibarra
committed
4508808

Commits on Aug 10, 2026

chore: merge main into B200 branch
adibarra
committed
f862b86
fix: test B200 with modern Kimi nightly
adibarra
committed
3bf7616
fix: disable fused Kimi latent tail
adibarra
committed
ef1ffac
fix: force portable Kimi collectives
adibarra
committed
eac9a89
fix: install Kimi compatibility setup
adibarra
committed
08ca238
fix: bypass unsupported Kimi workspaces
adibarra
committed
48ea8e6
fix: force PyNCCL for Kimi
adibarra
committed
6c86aff
fix: terminate Kimi changelog entry
adibarra
committed
0d8d71b
refactor: remove obsolete DEP accounting changes
adibarra
committed
cdcd4a7
fix: reserve sufficient Kimi offload memory
adibarra
committed
c4f8ad5
chore: remove stale DEP launcher comment
adibarra
committed
b37142c
fix: cap offload at stable concurrency
adibarra
committed
5167128
fix: retain accepted offload capacity endpoint
adibarra
committed
629d447

Commits on Aug 11, 2026

chore: merge main and resolve conflicts
adibarra
committed
4107dc6
fix: preserve changelog newline after merge
adibarra
committed
72b321e
chore: merge main and resolve conflicts [skip-sweep]
adibarra
committed
0a2c7eb
fix: use upstream kimi k3 b200 path
adibarra
committed
977edcc
fix: run b200 tep in eager mode
adibarra
committed
3be29f2
fix: keep b200 tensor parallel node local
adibarra
committed
7abf1db
chore: merge main and preserve changelog
adibarra
committed
02bed65
ci: shorten multi-node eval artifact names
adibarra
committed
a476441
fix: enforce Kimi K3 metrics collection
cquil11
committed
db438d5
perf: probe Kimi K3 offload crossover
cquil11
committed
2bb6eea

Commits on Aug 12, 2026