CollectiveX: kv-transfer suite — NIXL + MoRI-IO KV-cache handoff benchmark (stacked on #2489) - #2510

Open

Oseltamivir wants to merge 51 commits into

mainfrom

collectivex-kv-transfer

Commits on Aug 10, 2026

CollectiveX: add the kv-transfer suite (NIXL + MoRI-IO KV-cache handoff benchmark)
Oseltamivir
committed
6959aa3
CollectiveX: self-raise the memlock rlimit before KV pool registration
Oseltamivir
committed
2156e6d
CollectiveX: enable mi355x kv-transfer via MoRI-IO
Oseltamivir
committed
db04dc2
CollectiveX: pin KV legs to the operator's RDMA selectors
Oseltamivir
committed
91d25fc
CollectiveX: enable h200 kv-transfer via NIXL
Oseltamivir
committed
986c679
CollectiveX: add the mooncake backend and the mnnvl fabric lane
Oseltamivir
committed
3b84f54
CollectiveX: open the launcher identity gates to kv backends and validate gb-nv rdma legs
Oseltamivir
committed
cb3bf94
CollectiveX: retarget the kv-mla scenario to DeepSeek-V4-Pro's shape
Oseltamivir
committed
c3f1ef9
CollectiveX: transfer DeepSeek-V4-Pro's cache as vLLM allocates it, across batch sizes
Oseltamivir
committed
0db55e7
CollectiveX: keep sweep_matrix stdlib-only and give the test job numpy
Oseltamivir
committed
d7d3eb6
CollectiveX: give the sweep dispatch a suites selector
Oseltamivir
committed
fa156b4
CollectiveX: retry kv wheel installs with --break-system-packages
Oseltamivir
committed
c4b5a37
CollectiveX: pin mooncake off its NVLink-IPC transport on same-fabric pairs
Oseltamivir
committed
d2b3f78
CollectiveX: stop forcing mooncake onto MNNVL on gb-nv kv legs
Oseltamivir
committed
e84ec10
CollectiveX: sweep only the DeepSeek-V4-Pro shape
Oseltamivir
committed
a915197
CollectiveX: emit the kv artifact version as a number
Oseltamivir
committed
0e0cac1
CollectiveX: densify the kv batch axis and add a sampling trial
Oseltamivir
committed
fec40e3
CollectiveX: extend the kv ISL ladder to 8k-512k under a burst descriptor budget
Oseltamivir
committed
40d9203
CollectiveX: extend the kv batch ladder to 32
Oseltamivir
committed
7b93d7d
CollectiveX: batch ladder to 64 with a kv case guard on gb-nv
Oseltamivir
committed
4e9390c
CollectiveX: backfill-friendly allocation asks for kv legs
Oseltamivir
committed
f2f36c3
CollectiveX: mooncake on mi355x as a push-only row from AMD's atom-dev build
Oseltamivir
committed
11fe18d
CollectiveX: give mooncake's per-call guard headroom under contention
Oseltamivir
committed
7e788ce

Commits on Aug 12, 2026

Commits on Aug 27, 2026

Commits on Aug 28, 2026

Commits on Aug 29, 2026

CollectiveX: size the gb300 kv time budget to its measured pacing
Oseltamivir
committed
030a8c1

Commits on Aug 30, 2026

Commits on Aug 31, 2026