[c10d] Prototype of group_split for dist2 work (#157716)

This is to implement group_split as proposed in [docs.google.com/document/d/13R-1t_yESTvmAjcCN-wQjQQadIEu0JNIdS65uZawZzY/edit?tab=t.0#heading=h.3ctbqqopzc89](https://docs.google.com/document/d/13R-1t_yESTvmAjcCN-wQjQQadIEu0JNIdS65uZawZzY/edit?tab=t.0#heading=h.3ctbqqopzc89)

Pull Request resolved: https://github.com/pytorch/pytorch/pull/157716
Approved by: https://github.com/d4l3k
This commit is contained in:
fduwjj
2025-07-14 10:00:40 -07:00
committed by PyTorch MergeBot
parent 1e4d8b5a4a
commit 6b2bef10af
14 changed files with 246 additions and 7 deletions

View File

@ -231,6 +231,7 @@ static std::map<at::ScalarType, ncclDataType_t> ncclDataType = {
};
TORCH_API size_t hashTensors(const std::vector<at::Tensor>& tensors);
TORCH_API int genNcclSplitColor(const std::vector<int>& ranks);
TORCH_API std::string getNcclVersion();
TORCH_API std::tuple<int, int, int> getNcclVersionTuple();
TORCH_API int getNcclVersionNumber();