Two defects measured on 4x GB200, MiniMax-H3 geometry, per attention layer
(NCCL baseline 2421us at sp=4, 1295us at sp=8 across two trays):
as shipped sp=4 2035us (1.19x) sp=8 3102us (2.4x SLOWER than NCCL)
with these sp=4 1590us (1.52x) sp=8 declines, 1300us
ncclTeamLsa answers "addressable", not "fast". NCCL 2.29 extends the LSA
team across a multi-node NVLink domain, so on a GB200 rack the gate passes
for ranks on different trays and the kernel arms. Its fine-grained 16B
remote stores are far slower there than NCCL's bulk transfers. Require a
single host, which is the regime the slab decomposition was tuned for, and
which matches flashinfer's own gate. torch 2.12 (the pin) bundles NCCL
2.29.7, so this is reachable today.
_can_attempt now computes its local verdict without collectives and then
runs one unconditional all_gather_object carrying (hostname, local_ok).
A collective behind a rank-local early return hangs the group whenever
ranks disagree -- exactly the case this gate exists to detect. Verified:
with one rank reporting the kernel unavailable, the earlier ordering timed
out at 180s while this completes with correct results on every rank.
The per-call agreement cost a flat ~227us regardless of operand size -- a
host-side gloo all_gather before every collective. The signature is
architectural: two distinct values (scatter and gather shapes) across 50
layers x 4 steps. Cache the verdict so the collective runs twice per
generation instead of ~400 times.
The cache trades one property: a rank whose signature diverges mid-run now
misses the cache and calls the collective alone, hanging rather than
falling back. Re-voting every N calls would bound that if wanted.
shaoxiongduan
3eabb7b40b