feat(dspark): add Ulysses sequence parallel training - #728

Draft

maocheng23 wants to merge 26 commits into

sgl-project:mainfrom

maocheng23:feat/dspark-ulysses-sp

Commits on Jul 26, 2026

feat(dspark): add Ulysses sequence parallel training
maocheng23
committed
c72b08d

Commits on Jul 29, 2026

feat(dspark): support online Ulysses training
maocheng23
committed
b59040f
fix(disagg): reclaim long-lease captures after durable ack
maocheng23
committed
b708470
test(disagg): preserve ordinary remove compatibility
maocheng23
committed
dbb608b
fix(disagg): preserve launcher cleanup state
maocheng23
committed
4ebd072
fix disagg launcher rank isolation
maocheng23
committed
d450c9b
examples: reserve GLM capture memory
maocheng23
committed
10a3c01
examples: cap long-context target KV cache
maocheng23
committed
982ace2
docs: make GLM 120k memory settings page-exact
maocheng23
committed
ca8d008
examples: disable graphs for GLM 120k capture
maocheng23
committed
c1010a1
examples: disable overlap for GLM 120k capture
maocheng23
committed
e33ce2d
patch: use message queue for TP capture requests
maocheng23
committed
c29cf93
patch: anchor TP receiver hunk on v0.5.16
maocheng23
committed
d6031b1
examples: synchronize TP capture queue rounds
maocheng23
committed
f97332c
patch: keep native TP request broadcast
maocheng23
committed
c53e321
examples: skip synthetic warmup for 120k capture
maocheng23
committed
e83a6a5
patch: isolate TP capture request broadcast
maocheng23
committed
6a7a7d4
patch: await batched scheduler delivery
maocheng23
committed
95f493b
fix: use single-request capture transport
maocheng23
committed
7ad8740
patch: compress spec capture control requests
maocheng23
committed
4fdb5ae
fix: await single capture request delivery
maocheng23
committed
cf6c582
examples: expose long-context prefill limit
maocheng23
committed
9b90228
patch: trace capture IPC delivery
maocheng23
committed
2a9d4fa
fix: capture chunked long-context prefill
maocheng23
committed
6a5d61e
docs: size 120k online capture runtime
maocheng23
committed
afc271a
style: satisfy pre-commit hooks
maocheng23
committed
5ad56a1

Navigation Menu

Uh oh!

feat(dspark): add Ulysses sequence parallel training - #728

feat(dspark): add Ulysses sequence parallel training#728
maocheng23 wants to merge 26 commits into
sgl-project:mainfrom
maocheng23:feat/dspark-ulysses-sp

Commits on Jul 26, 2026

feat(dspark): add Ulysses sequence parallel training

Commits on Jul 29, 2026

feat(dspark): support online Ulysses training

fix(disagg): reclaim long-lease captures after durable ack

test(disagg): preserve ordinary remove compatibility

fix(disagg): preserve launcher cleanup state

fix disagg launcher rank isolation

examples: reserve GLM capture memory

examples: cap long-context target KV cache

docs: make GLM 120k memory settings page-exact

examples: disable graphs for GLM 120k capture

examples: disable overlap for GLM 120k capture

patch: use message queue for TP capture requests

patch: anchor TP receiver hunk on v0.5.16

examples: synchronize TP capture queue rounds

patch: keep native TP request broadcast

examples: skip synthetic warmup for 120k capture

patch: isolate TP capture request broadcast

patch: await batched scheduler delivery

fix: use single-request capture transport

patch: compress spec capture control requests

fix: await single capture request delivery

examples: expose long-context prefill limit

patch: trace capture IPC delivery

fix: capture chunked long-context prefill

docs: size 120k online capture runtime

style: satisfy pre-commit hooks