Description
With 635b98f.
As first noted in #132945 (comment), pytorch built with XPU backend is missing a range of ATen operator headers (missing installed: they are generated, but not actually installed). Specifically, only those headers are available which are generated from torch stock sources (i.e. from ./aten/src/ATen/native/native_functions.yaml), these are:
$ find torch/include/ATen/ops -name "*xpu*"
torch/include/ATen/ops/baddbmm_xpu_dispatch.h
torch/include/
806B
ATen/ops/bmm_xpu_dispatch.h
torch/include/ATen/ops/mm_xpu_dispatch.h
torch/include/ATen/ops/_addmm_activation_xpu_dispatch.h
torch/include/ATen/ops/addmm_xpu_dispatch.h
torch/include/ATen/ops/addmv_xpu_dispatch.h
torch/include/ATen/ops/addbmm_xpu_dispatch.h
For CUDA a lot more are generated (from the same .yaml file):
$ find torch/include/ATen/ops -name "*cuda*" | wc -l
604
For XPU however, those missing headers are generated separately from torch-xpu-ops. These are defined by https://github.com/intel/torch-xpu-ops/blob/main/yaml/native/native_functions.yaml. These XPU headers are generated, but not installed.
These missing headers might be used by some third party projects. On pytorch level these headers are used in some tests. For example, see #138088 which needs ATen/ops/cat_*_dispatch.h and ATen/ops/norm_*_dispatch.h.
Can missing aten xpu headers be installed?
I think that current challenge is that torch-xpu-ops generates not only XPU specific files from https://github.com/intel/torch-xpu-ops/blob/main/yaml/native/native_functions.yaml, but generic files as well. I.e. we get duplicated files generated from stock pytorch sources and from xpu sources. For example:
$ find . -name _aminmax.h
./build/aten/src/ATen/ops/_aminmax.h
./build/xpu/ATen/ops/_aminmax.h
./torch/include/ATen/ops/_aminmax.h
I guess there are 2 ways forward:
- Filter out non-XPU header files and install only them
- Move XPU headers generation to stock pytorch and drop XPU torch-xpu-ops level customization
CC: @gujinghui @EikanWang @fengyuan14 @guangyey @jgong5 @albanD
cc @seemethere @malfet @osalpekar @atalman @gujinghui @EikanWang @fengyuan14 @guangyey
Metadata
Metadata
Assignees
Labels
Type
Projects
Status