feat: support Qwen3.5 Decode Context Parallel on Ascend NPU. - #2125

Open

Enguikong wants to merge 28 commits into

xLLM-AI:mainfrom

Enguikong:feat/qwen35-dcp-correctness

Commits on Aug 5, 2026

Commits on Aug 6, 2026

feat: support DCP-aware chunked prefill
Enguikong
committed
def4450
merge: reconcile fork DCP branch history
Enguikong
committed
15e5217

Commits on Aug 10, 2026

fix: close Phase 4 DCP validation gaps
Enguikong
committed
1533da6
fix: fail closed DCP chunked prefill
Enguikong
committed
c80271c
fix: close mixed batch when decode context parallel size > 1

authored and
Super User
committed
7bac0b3
feat: gate DCP chunked prefill behind an experimental opt-in

authored and
Super User
committed
0de5541
fix: gate DCP chunked prefill on effective, not raw, chunked flag

authored and
Super User
committed
d7f86fe
fix: use effective chunked prefill in whole-service DCP consumers

authored and
Super User
committed
9f144b1
feat: allow Qwen3.5 MoE with decode context parallelism

authored and
Super User
committed
b319f2c
feat: warn on Qwen3.5 MoE with decode context parallelism

authored and
Super User
committed
05a8151

Commits on Aug 13, 2026

feat: allow schedule overlap with decode context parallelism

authored and
Super User
committed
04aee08

Commits on Aug 17, 2026

feat: support decode context parallelism with ACL graph
Enguikong
committed
3844bd5

Commits on Aug 20, 2026

Commits on Aug 24, 2026

Commits on Aug 25, 2026