Add B200 Qwen3.5 397B FP8 disaggregated AgentX recipes - #3269

Merged

cquil11 merged 36 commits into

mainfrom

config/b200-qwen35-fp8-dynamo-agentx

Sep 19, 2026

Commits on Sep 18, 2026

feat: add B200 Qwen3.5 FP8 disaggregated AgentX draft
cquil11
committed
10f1a60
docs: link B200 AgentX performance changelog to PR
cquil11
committed
aca85d8
feat: align B200 disaggregated AgentX with aggregate baseline
cquil11
committed
33371d2
perf: use latest SGLang nightly for B200 Qwen3.5 FP8
cquil11
committed
e64b9c1
Merge remote-tracking branch 'origin/main' into config/b200-qwen35-fp8-dynamo-agentx
cquil11
committed
134479b
fix: align Dynamo with latest SGLang module layout
cquil11
committed
d1368f7
fix: handle B200 Qwen3.5 node-local weights in SRT preflight
cquil11
committed
683c6c8
fix: use intra-node NVLink for colocated B200 Qwen3.5 FP8
cquil11
committed
64234cd
Merge main and retain Qwen FP8 node-local model handling
cquil11
committed
e214038
perf: add isolated latest B200 Qwen3.5 FP8 aggregate candidates
cquil11
committed
f021964
fix: preserve standalone B200 aggregate budget defaults
cquil11
committed
b3d72d1
fix: propagate failed B200 Slurm sweeps after artifact collection
cquil11
committed
e82ee03
fix: use verified NVIDIA wheels for B200 Dynamo dependencies
cquil11
committed
b03becb
feat: select measured B200 FP8 HiCache disaggregation candidate
cquil11
committed
2560380
feat: qualify B200 FP8 aggregate high-concurrency frontier candidate
cquil11
committed
58fe416
test: include runtime inputs in isolated B200 launcher fixture
cquil11
committed
5fd1bed
feat: stage B200 FP8 2P1D frontier candidate for qualification
cquil11
committed
3dcbfda
fix: require explicit aggregate HiCache budget mode from caller
cquil11
committed
27b44dc
merge: integrate current main before expanded B200 qualification
cquil11
committed
a591c4e
fix: preserve shared log streaming while verifying B200 job status
cquil11
committed
873bf04
fix: preserve published aggregate curves in B200 qualification scope
cquil11
committed
6498b90
feat: expand B200 FP8 frontier qualification with concurrency 16
cquil11
committed
1efe0d3
feat: select B200 FP8 concurrency 24 over dominated 2P1D candidate
cquil11
committed
8170706
feat: qualify B200 concurrency 8 against canonical aggregate evidence
cquil11
committed
a83fe81
Merge remote-tracking branch 'origin/main' into config/b200-qwen35-fp8-dynamo-agentx
cquil11
committed
f3b787e
fix: require validated B200 Qwen3.5 AgentX power telemetry
cquil11
committed
cbb19c6
feat: qualify B200 FP8 concurrency 64 with first-use HiCache backup
cquil11
committed
6a9ecdb
fix: supply B200 AgentX power window contract
cquil11
committed
29d5097
chore: sync B200 frontier recipes with main
cquil11
committed
9372a1b
config: add measured B200 FP8 C48 write-through point
cquil11
committed
beae237
chore: sync B200 recipes with latest image updates
cquil11
committed
c4c710e
docs: consolidate B200 PR performance changelog
cquil11
committed
0b75e13

Commits on Sep 19, 2026