Repository navigation

[PyTorch] Share CUDA graph memory across dynamic CP variants - #3353

Open

xiaoyao0115 wants to merge 6 commits into

NVIDIA:mainfrom

xiaoyao0115:test/dcp-cg-gap-fix

Open

[PyTorch] Share CUDA graph memory across dynamic CP variants#3353
xiaoyao0115 wants to merge 6 commits into
NVIDIA:mainfrom
xiaoyao0115:test/dcp-cg-gap-fix

Commits on Nov 11, 2026
- [PyTorch] Share CUDA graph memory across graph variants with slot arenas
  
  xiaoyao0115
  and
  claude
  committed
  83d11ca
- Protect slot graph outputs and captured attention state
  
  xiaoyao0115
  and
  claude
  committed
  485b1a8
- Snapshot saved activations before asynchronous graph offload
  
  xiaoyao0115
  and
  claude
  committed
  9fc2372
- Keep the two-step THD Q metadata layout for NO_LOAD_BALANCE all-gather CP
  
  xiaoyao0115
  and
  claude
  committed
  58c3715
- Require slot branches to share arenas and document async offload copies
  
  xiaoyao0115
  and
  claude
  committed
  ec656d3
- Check layouts of module-state outputs across slot branches
  
  xiaoyao0115
  and
  claude
  committed
  15fe37e