fix(pd): harden production cutover lifecycles - #1472

Closed

sufubao wants to merge 23 commits into

ModelTC:mainfrom

sufubao:deploy/production-cutover-readiness-20260819

Commits on Aug 13, 2026

fix(pd): avoid first-token cache stats race
sufubao
committed
928d6bd
fix(pd): preserve prefill metadata and unblock decode
sufubao
committed
e71f84a
refactor(pd): simplify first-token fallback
sufubao
committed
c912a63
chore: keep PD fix PR focused
sufubao
committed
b05ff81
refactor(pd): consume node mode once
sufubao
committed
00f2698
refactor(pd): clarify first-token buffering
sufubao
committed
e4604cb
refactor(pd): make prefill lookup explicit
sufubao
committed
98d8626
refactor(pd): avoid rescanning buffered tokens
sufubao
committed
29ce20c
refactor(pd): clarify first-token stream handling
sufubao
committed
c979dd0
refactor(pd): streamline first-token buffering
sufubao
committed
750a02e
refactor(pd): reuse async queue for output tokens
sufubao
committed
563e4dd
style(pd): apply project formatting
sufubao
committed
aabe3b6
fix(pd): bound missing prefill token fallback
sufubao
committed
0f56003
Delete unit_tests/server/httpserver/test_pd_master_cached_tokens.py
sufubao
committed
14448c3
fix(pd): preserve terminal decode marker
sufubao
committed
5923736

Commits on Aug 19, 2026

Commits on Aug 20, 2026

fix(pd): cancel requests waiting for decode slots
sufubao
committed
348aa88
fix(pd): fail requests when assigned nodes disconnect
sufubao
committed
f6a4417
fix(pd): propagate worker background failures
sufubao
committed
623d640
fix(pd): recover NIXL transfer task failures
sufubao
committed
feedc43
style(pd): apply repository pre-commit formatting
sufubao
committed
a4172be