Add support for the operator 'aten::__lshift__.Scalar' for the XPU device

[ZzEeKkAa]

🚀 The feature, motivation and pitch

I'm getting this error while running intel's triton unit tests with upstream pytorch:

NotImplementedError: The operator 'aten::lshift.Scalar' is not currently implemented for the XPU device. Please open a feature on https://github.com/intel/torch-xpu-ops/issues. You can set the environment variable PYTORCH_ENABLE_XPU_FALLBACK=1 to use the CPU implementation as a fallback for XPU unimplemented operators. WARNING: this will bring unexpected performance compared with running natively on XPU.

Alternatives

No response

Additional context

No response

[fengyuan14]

Relatives,

• [ ] aten::lshift
• [ ] aten::ilshift
• [ ] aten::rshift
• [ ] aten::irshift

[anmyachev]

Hi @fengyuan14,

since __rshift__.Scalar is already in the list of operations that fall back to the CPU, does it make sense to add __lshift__.Scalar and relatives to that list as well to speed up getting a working, albeit slow, version of these operations? https://github.com/intel/torch-xpu-ops/blob/35bea25e2e09b92067349a75cd0858a485c444fe/src/ATen/native/xpu/XPUFallback.template#L243

[ZzEeKkAa]

@anmyachev I've just checked your proposal and it is indeed resolves the problem. However native support is proffered it is just a working solution

[fengyuan14]

Hi @fengyuan14,

since __rshift__.Scalar is already in the list of operations that fall back to the CPU, does it make sense to add __lshift__.Scalar and relatives to that list as well to speed up getting a working, albeit slow, version of these operations?

https://github.com/intel/torch-xpu-ops/blob/35bea25e2e09b92067349a75cd0858a485c444fe/src/ATen/native/xpu/XPUFallback.template#L243

Hi, please let me clarify our consideration. The operator coverage plan of PT2.4/PT2.5 is decided by PyTorch usages what we commit to support. The usage scope is to support,

1. PyTorch Dynamo benchmark, which includes part of models in HF/TIMM/TB.
2. Operators MPS supports.

The operators in the issue are not listed in our existing plan. However, the plan is changeable, we are listening to the community if there is any urgent requirement. And we will balance our limited efforts and pick them on at a proper moment. Thanks for your inputs. For the operators listed here, we will try to catch up PT2.5.

[anmyachev]

@fengyuan14 thank you for such a detailed and quick response!

[ZzEeKkAa]

Closing it, since it was implemented in #688