Repository navigation

[None][feat] Staircase: Rubin GLM Optimized Low Latency (Conc up to 56) - #21380

Merged

dongfengy merged 15 commits into

NVIDIA:mainfrom

dongfengy:feat/glm-rubin-staircase-four-points

Nov 6, 2026

Merged

[None][feat] Staircase: Rubin GLM Optimized Low Latency (Conc up to 56)#21380
dongfengy merged 15 commits into
NVIDIA:mainfrom
dongfengy:feat/glm-rubin-staircase-four-points

Commits on Nov 5, 2026
- feat: extend Rubin GLM low-latency targets within the GLM family
  dongfengy
  committed
  8084d07
- fix: wait before loading Rubin IPC residuals
  dongfengy
  committed
  66683a1
- fix: use None for ready Rubin MoE inputs
  dongfengy
  committed
  2912aa3
- fix: require 32 indexer heads for Rubin GLM routing
  dongfengy
  committed
  1cf498f
- test: register Rubin small-batch MTP CPU coverage
  dongfengy
  committed
  d43fd26
- test: guard Rubin IPC matrix GPU count
  dongfengy
  committed
  a797250
- test: register four Rubin accuracy cases for QA
  dongfengy
  committed
  672677f
- test: register four-rank Rubin op matrices for QA
  dongfengy
  committed
  56a373a
- fix: clarify Rubin sampling validation message
  dongfengy
  committed
  f075019
- fix: log Rubin draft eligibility failures
  dongfengy
  committed
  8aacd58
- test: refresh latent MLA inputs between graph replays
  dongfengy
  committed
  0d3c533
- test: cover Rubin FC2 token-group boundaries
  dongfengy
  committed
  40db791
- fix: follow the shared all-reduce cache for Rubin IPC
  dongfengy
  committed
  047240a
- fix: keep Rubin IPC clear lengths in bytes
  dongfengy
  committed
  a1cd54b
- fix: complete Rubin binding license headers
  dongfengy
  committed
  0e5005d