feat(dsv4): add B200 DeepSeek-V4-Pro dynamo-trt 8k1k disagg throughput recipes / 新增 B200 DeepSeek-V4-Pro dynamo-trt 8k1k 分离式吞吐量配置 - #2721

Open

richardhuo-nv wants to merge 16 commits into

mainfrom

rihuo/dsv4-b200-dynamo-trt-8k1k

Commits on Aug 24, 2026

feat(dsv4): add B200 DeepSeek-V4-Pro dynamo-trt 8k1k disagg throughput recipes

richardhuo-nv
and
claude
committed
eab149a
changelog(dsv4-b200): set pr-link to #2721

richardhuo-nv
and
claude
committed
4604d43
refactor(dsv4-b200): split 8k1k/b200 dynamo-trt recipes into mtp/ and stp/

richardhuo-nv
and
claude
committed
bef13cc
chore(dsv4-b200): drop RESULTS.md/png, fold split into one changelog entry

richardhuo-nv
and
claude
committed
807a2c6
refactor(dsv4-b200): co-locate moe_load_balancer configs under 8k1k/b200

richardhuo-nv
and
claude
committed
6d2fcd9
fix(dsv4-b200): pin srt-slurm to v1.0.69 for dynamo-trt

richardhuo-nv
and
claude
committed
73f08a9
add new line
richardhuo-nv
committed
04895a4
fix(configs): drop stray ep/dp-attn lines before dsv4-fp4-b200-dynamo-trt

richardhuo-nv
and
claude
committed
14289a7

Commits on Aug 25, 2026

fix(dsv4-b200): strip forced MTP acceptance length on EVAL_ONLY runs

richardhuo-nv
and
claude
committed
1e80df9
chore(dsv4-b200): bump to trtllm 1.3.0rc24 and dynamo wheel 1.4.0.dev20260807

richardhuo-nv
and
claude
committed
d0026a9
Merge branch 'main' into rihuo/dsv4-b200-dynamo-trt-8k1k

richardhuo-nv
and
claude
committed
577442f
changelog(dsv4-b200): drop upstream PR reference from the entry

richardhuo-nv
and
claude
committed
0aafb03
perf(dsv4-b200): retune 8k1k concurrencies and sync master conc-list

richardhuo-nv
and
claude
committed
b31bf0e
boost tokenizer
richardhuo-nv
committed
ba30f69
Merge branch 'main' into rihuo/dsv4-b200-dynamo-trt-8k1k
richardhuo-nv
committed
e15f64d

Commits on Aug 26, 2026

disable thinking
richardhuo-nv
committed
09804dc