[None][refactor] overhaul NVLink one-sided all-to-all - #19610

Open

zhangcl wants to merge 26 commits into

NVIDIA:mainfrom

zhangcl:user/chulianz/nvlink-one-sided-overhaul

Commits on Sep 25, 2026

[None][fix] make CUDA graph kernel profiling CFT-compatible

bobboli
authored and
zhangcl
committed
ffb177e
[None][fix] add CFT driver-version detection helpers
zhangcl
committed
643f616
[None][fix] enable automatic CFT selection and CUPTI event profiling

bobboli
authored and
zhangcl
committed
e026c5c
[None][refactor] simplify MoE communication benchmark and add model profiles

bobboli
authored and
zhangcl
committed
2917c1b
[None][refactor] namespace one-sided A2A controls and fix block sizes

bobboli
authored and
zhangcl
committed
b34e46d
[None][refactor] isolate one-sided A2A workspace regions and add round-trip tests

bobboli
authored and
zhangcl
committed
c80d983
[None][refactor] consolidate NVLink one-sided communication tests

bobboli
authored and
zhangcl
committed
ca50319
[None][refactor] separate MNNVL memory from two-sided MoE communication

bobboli
authored and
zhangcl
committed
8759219
[None][test] consolidate one-sided CFT policy coverage

bobboli
authored and
zhangcl
committed
bafb66d
[None][perf] use CUPTI kernel spans for MoE communication benchmarks

bobboli
authored and
zhangcl
committed
fc13ffa
[None][refactor] select NVLink one-sided A2A timeouts in Python

bobboli
authored and
zhangcl
committed
4fd8715
[None][refactor] tidy CFT kernel naming and synchronization

bobboli
authored and
zhangcl
committed
b663bb0
[None][fix] guard CFT dispatch workspace reuse across rounds

bobboli
authored and
zhangcl
committed
7b48014
[None][perf] compact NVLink one-sided dispatch and combine fanout

bobboli
authored and
zhangcl
committed
f602a13
[None][refactor] share NVLink one-sided round flag helpers

bobboli
authored and
zhangcl
committed
7fbd2cb
[None][fix] skip invalid expert routes in NVLink one-sided dispatch

bobboli
authored and
zhangcl
committed
8ad4d24
[None][fix] reconcile the one-sided overhaul with main-only callers
zhangcl
committed
23b2799
[None][fix] drop the Python combine-offset reservation superseded by native regions
zhangcl
committed
2f1d9cd
[None][refactor] simplify NVLink one-sided dispatch and combine

bobboli
authored and
zhangcl
committed
750294d
[None][refactor] organize NVLink one-sided workspace by phase

bobboli
authored and
zhangcl
committed
130dc98
[None][fix] re-initialize the planned one-sided layout on MNNVL restore
zhangcl
committed
3640eff
[None][perf] pair FP8 conversions in one-sided combine

bobboli
authored and
zhangcl
committed
3737131
[None][docs] clarify one-sided workspace metadata fields

bobboli
authored and
zhangcl
committed
7eeb699
[None][fix] use main's MoE A2A paths in ported sources
zhangcl
committed
a7695a0
[None][fix] honor an explicit CFT choice when sizing the one-sided workspace
zhangcl
committed
0d9a563
[None][fix] import NVLinkOneSided from main's MoE package in model_engine
zhangcl
committed
6136b7a

Navigation Menu

[None][refactor] overhaul NVLink one-sided all-to-all - #19610

[None][refactor] overhaul NVLink one-sided all-to-all#19610
zhangcl wants to merge 26 commits into
NVIDIA:mainfrom
zhangcl:user/chulianz/nvlink-one-sided-overhaul

Commits on Sep 25, 2026

[None][fix] make CUDA graph kernel profiling CFT-compatible

[None][fix] add CFT driver-version detection helpers

[None][fix] enable automatic CFT selection and CUPTI event profiling

[None][refactor] simplify MoE communication benchmark and add model profiles

[None][refactor] namespace one-sided A2A controls and fix block sizes

[None][refactor] isolate one-sided A2A workspace regions and add round-trip tests

[None][refactor] consolidate NVLink one-sided communication tests

[None][refactor] separate MNNVL memory from two-sided MoE communication

[None][test] consolidate one-sided CFT policy coverage

[None][perf] use CUPTI kernel spans for MoE communication benchmarks

[None][refactor] select NVLink one-sided A2A timeouts in Python

[None][refactor] tidy CFT kernel naming and synchronization

[None][fix] guard CFT dispatch workspace reuse across rounds

[None][perf] compact NVLink one-sided dispatch and combine fanout

[None][refactor] share NVLink one-sided round flag helpers

[None][fix] skip invalid expert routes in NVLink one-sided dispatch

[None][fix] reconcile the one-sided overhaul with main-only callers

[None][fix] drop the Python combine-offset reservation superseded by native regions

[None][refactor] simplify NVLink one-sided dispatch and combine

[None][refactor] organize NVLink one-sided workspace by phase

[None][fix] re-initialize the planned one-sided layout on MNNVL restore

[None][perf] pair FP8 conversions in one-sided combine

[None][docs] clarify one-sided workspace metadata fields

[None][fix] use main's MoE A2A paths in ported sources

[None][fix] honor an explicit CFT choice when sizing the one-sided workspace

[None][fix] import NVLinkOneSided from main's MoE package in model_engine