[TRTLLM-16304][feat] In-tree implementation of staircase - #19056

Merged

WeiHaocheng merged 19 commits into

NVIDIA:mainfrom

WeiHaocheng:feat/staircase-upstream

Sep 27, 2026

Commits on Sep 27, 2026

[TRTLLM-16304][feat] In-tree implementation of staircase
WeiHaocheng
committed
18d47b7
[TRTLLM-16304][test] Wire the staircase catalog tests into the unit test suite
WeiHaocheng
committed
8faab78
[TRTLLM-16304][test] Add the staircase whole-model gates to the accuracy suite
WeiHaocheng
committed
4175d7a
[TRTLLM-16304][test] Keep the staircase collectives off xdist, drop a redundant gate
WeiHaocheng
committed
c9fb814
[TRTLLM-16304][test] Rename the staircase collective rank bodies off the test naming
WeiHaocheng
committed
a566f1b
[TRTLLM-16304][chore] Address the in-tree migration review
WeiHaocheng
committed
b52ec33
[TRTLLM-16304][chore] Rename staircase to modeling_v2
WeiHaocheng
committed
16a5e07
[TRTLLM-16304][chore] Follow upstream's removal of MTPDraftModelForCausalLM
WeiHaocheng
committed
62c771c
[TRTLLM-16304][fix] Certify thop_attention over two KV pools
WeiHaocheng
committed
48dec0d
[TRTLLM-16304][infra] Add a CBTS rule for modeling_v2
WeiHaocheng
committed
2d82d7a
[TRTLLM-16304][test] Cut the catalog tests down to what the targets run
WeiHaocheng
committed
6ae72b5
[TRTLLM-16304][chore] Move modeling_v2 under _experimental
WeiHaocheng
committed
f9b71e8
[TRTLLM-16304][chore] Fix the codespell finding blocking pre-commit
WeiHaocheng
committed
d97d768
[TRTLLM-16304][test] Carry the modeling_v2 switch on the case, not the shell
WeiHaocheng
committed
4c9b605
[TRTLLM-16304][test] Move the single-GPU modeling_v2 entries off the 4-GPU stage
WeiHaocheng
committed
d43692f
[TRTLLM-16304][test] Put the single-GPU modeling_v2 entries on a stage that accepts them
WeiHaocheng
committed
ff84247
[TRTLLM-16304][test] Take the prose back out of the test-db lists
WeiHaocheng
committed
652be0a
[TRTLLM-16304][chore] Flatten a target's path into one directory name
WeiHaocheng
committed
e1c23f4
[TRTLLM-16304][chore] Rewrap two lines the flattened path made short enough to join
WeiHaocheng
committed
8622232

Navigation Menu

[TRTLLM-16304][feat] In-tree implementation of staircase - #19056

[TRTLLM-16304][feat] In-tree implementation of staircase#19056
WeiHaocheng merged 19 commits into
NVIDIA:mainfrom
WeiHaocheng:feat/staircase-upstream

Commits on Sep 27, 2026

[TRTLLM-16304][feat] In-tree implementation of staircase

[TRTLLM-16304][test] Wire the staircase catalog tests into the unit test suite

[TRTLLM-16304][test] Add the staircase whole-model gates to the accuracy suite

[TRTLLM-16304][test] Keep the staircase collectives off xdist, drop a redundant gate

[TRTLLM-16304][test] Rename the staircase collective rank bodies off the test naming

[TRTLLM-16304][chore] Address the in-tree migration review

[TRTLLM-16304][chore] Rename staircase to modeling_v2

[TRTLLM-16304][chore] Follow upstream's removal of MTPDraftModelForCausalLM

[TRTLLM-16304][fix] Certify thop_attention over two KV pools

[TRTLLM-16304][infra] Add a CBTS rule for modeling_v2

[TRTLLM-16304][test] Cut the catalog tests down to what the targets run

[TRTLLM-16304][chore] Move modeling_v2 under _experimental

[TRTLLM-16304][chore] Fix the codespell finding blocking pre-commit

[TRTLLM-16304][test] Carry the modeling_v2 switch on the case, not the shell

[TRTLLM-16304][test] Move the single-GPU modeling_v2 entries off the 4-GPU stage

[TRTLLM-16304][test] Put the single-GPU modeling_v2 entries on a stage that accepts them

[TRTLLM-16304][test] Take the prose back out of the test-db lists

[TRTLLM-16304][chore] Flatten a target's path into one directory name

[TRTLLM-16304][chore] Rewrap two lines the flattened path made short enough to join