Repository navigation

[PyT] [Common] add support for enabling cuda graph under thd format in megatron. - #2898

Merged

timmoon10 merged 12 commits into

NVIDIA:mainfrom

HaochenYuan:thd_cuda_graph_0415

Jul 1, 2026

Merged

[PyT] [Common] add support for enabling cuda graph under thd format in megatron.#2898
timmoon10 merged 12 commits into
NVIDIA:mainfrom
HaochenYuan:thd_cuda_graph_0415

Commits on Jun 5, 2026
- add support for THD CUDA graph
  HaochenYuan
  committed
  2c58118
Commits on Jun 6, 2026
- modify comment
  HaochenYuan
  committed
  b6aa597
Commits on Jun 8, 2026
- Merge branch 'main' into thd_cuda_graph_0415
  HaochenYuan
  authored
  69fddf7
Commits on Jun 9, 2026
- address @timmoon10: drop FAv2-bwd alloc gate, rely on THD tail zero-fill
  HaochenYuan
  committed
  805746a
Commits on Jun 11, 2026
- Support graph-safe MoE aux loss token count
  HaochenYuan
  committed
  a2cc7ef
- add graph guard for one zero fill
  HaochenYuan
  committed
  21d8d3a
- remove redundant zero-fill
  HaochenYuan
  committed
  cee7b6a
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  pre-commit-ci[bot]
  committed
  823805d
Commits on Jun 30, 2026
- rename & remove prelude kernel
  HaochenYuan
  committed
  bce6bc5
- Merge branch 'main' into thd_cuda_graph_0415
  timmoon10
  authored
  edfdb78
- Update warp reduction function
  timmoon10
  authored
  a69fb1b
- [pre-commit.ci] auto fixes from pre-commit.com hooks
  pre-commit-ci[bot]
  committed
  1bc41e5