Add B300 Qwen3.5 397B FP8 disaggregated AgentX recipes - #3268

Open

cquil11 wants to merge 40 commits into

mainfrom

cam/qwen35-b300-disagg

Commits on Sep 18, 2026

feat: add B300 Qwen3.5 FP8 disaggregated AgentX starter
cquil11
committed
ad1e960
chore: link B300 starter performance changelog
cquil11
committed
1803e0d
perf: colocate B300 Qwen3.5 prefill and decode workers
cquil11
committed
d8cd1b4
Merge main into B300 Qwen3.5 AgentX draft
cquil11
committed
ca03c30
fix: match B300 Mamba cache flag to pinned SGLang
cquil11
committed
c66edd3
perf: use latest SGLang nightly for B300 Qwen3.5 FP8
cquil11
committed
347d6d0
fix: use Dynamo compatibility with latest SGLang
cquil11
committed
188f177
fix: use intra-node NVLink for colocated B300 PD
cquil11
committed
a945910
perf: select measured B300 FP8 HiCache PD candidate
cquil11
committed
f882301
fix: propagate failed B300 Slurm runs after collecting artifacts
cquil11
committed
fa54204
Merge main for expanded B300 qualification
cquil11
committed
7356847
Prepare expanded measured B300 FP8 qualification curve
cquil11
committed
08c5597
perf(b300): qualify measured concurrency 48 PD point
cquil11
committed
034155e
perf(b300): preserve existing aggregate publication scopes
cquil11
committed
ffa5ca1
style(b300): normalize final config newline
cquil11
committed
3f1f966
fix: require validated B300 Qwen3.5 AgentX power telemetry
cquil11
committed
06749eb
fix: bind B300 power windows to matrix concurrency
cquil11
committed
ebb3b09
perf: qualify B300 FP8 write-through PD at concurrency 96
cquil11
committed
434c5be
style: format B300 power contract fixture
cquil11
committed
ddc3ac1
fix: declare the B300 AgentX formal power window contract
cquil11
committed
0b07767
Merge latest main before B300 FP8 final qualification
cquil11
committed
62c951a
docs: consolidate the B300 FP8 performance changelog
cquil11
committed
a9e1e3e
Merge main before the expanded B300 FP8 sweep
cquil11
committed
1f5e468
Avoid unverified host-memory pressure for B300 H298 recipes
cquil11
committed
88b0e1e

Commits on Sep 19, 2026

chore: sync main while preserving the B300 sweep commit
cquil11
committed
dcf6661
fix: allow complete B300 AgentX runs on eligible nodes
cquil11
committed
a55052c
perf: retain B300 FP8 candidates supported by canonical qualification
cquil11
committed
48a9c30
chore: prepare B300 FP8 post-sweep main synchronization
cquil11
committed
4890295
chore: sync B300 frontier preparation with main
cquil11
committed
22e6784
feat(models): acquire verified original BF16 MTP shards atomically
cquil11
committed
0bbd8fa
test(models): reject corrupted provenance and unexpected asset files
cquil11
committed
4992470
feat(b300): stage original BF16 MTP assets for Qwen FP8 recipes
cquil11
committed
8ba2522
feat(b300): prepare C48 disaggregation with 64 decode slots
cquil11
committed
b73f488
perf(b300): add screened TP2 EP2 write-through candidate
cquil11
committed
b986edd
perf(b300): select measured C24 decode capacity
cquil11
committed
481e1ac
perf: expand B300 FP8 disaggregated AgentX frontier
cquil11
committed
3d6c4fe
fix: allow full B300 TP2 sweep startup and drain time
cquil11
committed
331938b
perf: select B300 1P2D throughput frontier point
cquil11
committed
6971261
perf: select B300 1P2D C32 frontier point
cquil11
committed
ed6d081
perf: add B300 1P2D C64 frontier point
cquil11
committed
b5f72ca