search

NVIDIA / Fuser

A Fusion Code Generator for NVIDIA GPUs (commonly known as "nvFuser")
Other
271 stars 53 forks source link

Add ReductionParams to python frontend #3431

Closed rdspring1 closed 2 days ago

rdspring1 commented 3 days ago

This PR add support for ReductionParams to the python frontend. All parameters are exposed except std::vector<TensorView*> smem_persistent_buffers and CompileParams.

liqiangxl commented 3 days ago

why CompileParams is skipped? When generate heuristic, outer reduction and all persistent schedulders will set the maxrregcount in compiler params.

rdspring1 commented 3 days ago

why CompileParams is skipped?

The python class is a reference to the original CPP class. It exists but just isn't exposed to the user. It is to be added later.

rdspring1 commented 3 days ago

!test