saymrwulf/pytorch
1
0
Fork 0
mirror of https://github.com/saymrwulf/pytorch.git synced 2026-05-14 20:57:59 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

[doc] fix grammar in "Extending Torch" (#140209)

Browse source 
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140209
Approved by: https://github.com/soulitzer
This commit is contained in:
Tongzhou Wang 2024-11-13 05:34:41 +00:00 committed by PyTorch MergeBot
parent 1886e33f60
commit 7b0d199471
1 changed files with 1 additions and 1 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
docs/source/notes/extending.rst
View file

@ -905,7 +905,7 @@ Some important implications of this implementation are:
- When calling back to Python and when wrapping the results, the same conversions are used as the regular PyTorch Python/C++ binding. In particular, some objects cannot be represented in Python and need special handling (undefined Tensors for example become None).
- Our native functions are lazily populated as ``torch.ops.{namespace}.{func_name}.{overload_name}`` as callable Python objects to enable easily interacting with them from Python. The ``func`` object given to ``__torch_dispatch__`` is always an entry from this namespace. This namespace can be used to directly call native ops and bypass the usual Python API and binding code.
In a similar way where ``__torch_function__`` is able to interpose on all of torch's Python API and Tensor methods, ``__torch_dispatch__`` is able intercepting all calls into the aten native API. Note that all methods on Tensors are converted into function calls before entering the dispatcher and thus will appear as function calls here: ``torch.add(a, 2)`` and ``a + 2`` will lead to exactly the same aten call.
In a similar way where ``__torch_function__`` is able to interpose on all of torch's Python API and Tensor methods, ``__torch_dispatch__`` is able to intercept all calls into the aten native API. Note that all methods on Tensors are converted into function calls before entering the dispatcher and thus will appear as function calls here: ``torch.add(a, 2)`` and ``a + 2`` will lead to exactly the same aten call.
Most of these functions are defined in ``native_functions.yaml`` which specifies the properties of these functions as well as their backend implementation. Their implementation alongside specified features are then automatically registered via codegen.
Some more exotic functions or features are also registered in other places in the C++ codebase or in user-defined C++ extensions.