[AMD-yanfeiwang]

Scope

Defensive correctness fix for a latent TP-collective desynchronization in the
decode-side hierarchical cache path. It only affects deployments that run
decode with --enable-hierarchical-cache under TP > 1 (where
decode_hicache_mixin calls is_load_back_event_done). Configs that do not
enable decode-side hierarchical cache never reach this code path, so this is a
no-op for them.

Problem

HiRadixCache.is_load_back_event_done() conditionally calls loading_check()
— which issues a TP all_reduce — only on ranks whose local load-back
finish_event has already fired:

if not finish_event.query():
    return False
self.loading_check()   # TP all_reduce, runs on a per-rank-divergent subset
return True

This helper is invoked per request from the decode load-back path
(disaggregation/decode_hicache_mixin.py), so both the number of calls and the
local finish_event.query() result differ across TP ranks. The conditional
collective therefore desynchronizes the TP all_reduce sequence and can
deadlock under TP > 1 (manifesting as a few requests stuck forever with
errors=0 and idle GPUs during high-concurrency decode-side HiCache load-back).

Fix

Make is_load_back_event_done() a pure per-rank GPU-event poll
(return finish_event.query()) and drop the collective. The load ack-reaping
collective is already issued exactly once per scheduler iteration at a
rank-invariant point (check_hicache_events -> loading_check), so acks are
still reaped in lockstep; this helper only needs to report local readiness for
per-request scheduling. This mirrors the existing rank-invariant handling in
writing_check().

Notes / test status

• python -m py_compile passes.
• This is a reasoned correctness hardening: it removes a per-rank-conditional TP
  collective, which is unsafe by construction. It has not been independently
  bisected against a live decode-hierarchical-cache TP>1 deadlock repro (our
  observed high-concurrency warmup hang was a separate router/PD-dispatch issue),
  so review of the lockstep-reaping argument above is appreciated.
• CI

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[github-actions]

Thanks @AMD-yanfeiwang. Closing this because it has had no updates in 91 days.

Reopen it if the work is still relevant.

Some directories moved recently, so an older branch may need retargeting:
sgl-kernel/ -> python/sglang/kernels/aot/, python/sglang/jit_kernel/
-> python/sglang/kernels/jit/, docs/ -> docs/docs/ (.mdx),
bench_serving.py -> benchmark/serving.py, test/srt/ -> test/registered/.