[None][feat] Rubin K3 merge-back cleanup: CFT release op, MegaMoE multi-node bootstrap, DeepGEMM JIT warmup - #19183

Open

reasonsolo wants to merge 11 commits into

NVIDIA:mainfrom

reasonsolo:user/lizhiz/mb-moe

Commits on Sep 22, 2026

[None][feat] MoE / MegaMoE Rubin (SM107) support
reasonsolo
committed
9a16d6d
[None][fix] Restore the moe_a2a_cft_release op
reasonsolo
committed
c056fb2
[None][fix] bind MegaMoE ranks to their local CUDA device
reasonsolo
committed
1a9e3f2
[None][fix] make MoE teardown and MASTER_PORT reuse fail cleanly
reasonsolo
committed
c1142ad
[None][fix] mask the Rubin manual-MMA scale-factor descriptor fields
reasonsolo
committed
8b8169b
[None][fix] cap MoE tactic profiling at the reachable token count
reasonsolo
committed
4458a87
[None][chore] Core: deepgemm JIT warmup, postproc metrics, cleanup
reasonsolo
committed
e05415a
[None][test] Port the disagg and KV-cache-v2 regression tests from the Rubin branch
reasonsolo
committed
ace591d
[None][fix] restore the V2 draft-KV max_tokens derivation dropped in the rebase
reasonsolo
committed
27cd6cf
[None][test] restore the MegaMoE local-device bootstrap test
reasonsolo
committed
ffdf3f3
[None][fix] keep the KDA dtype cast usable under meta init
reasonsolo
committed
d8f6fa6