fb9a5d248f
Fixes #141563 In NumPy, an ellipsis always acts as a separator between advanced indices, even when the ellipsis doesn't actually match any dimensions. In PyTorch an empty ellipsis doesn't cause a separation. This leads to differing behavior between Numpy and PyTorch in this edge case. This difference in behavior leads to a bug when using torch.compile: ```python >>> import numpy as np >>> f = lambda x: x[:,(0,1),...,(0,1)].shape >>> a = np.ones((3, 4, 5)) >>> f(a) (2, 3) >>> torch.compile(f)(a) (3, 2) ``` Similarly to #157676, this PR doesn't change PyTorch's behavior, but it fixes the translation layer, ensuring torch._numpy compatibility with NumPy. I am marking this PR as fixing #141563, even though PyTorch behavior isn't modified. Notice that there are still some other bugs in PyTorch's advanced indexing, that need to be fixed (mainly regarding proper accounting of dimensions when multidimensional boolean masks are present). But those need to be fixed at the ATen operator level. Examples: - #71673 - #107699 - #158125 Pull Request resolved: https://github.com/pytorch/pytorch/pull/158297 Approved by: https://github.com/soumith
NumPy <> PyTorch Compat Layer
This folder contains an implementation of (most of) the NumPy public API using PyTorch tensors. Note that this folder does not depend on NumPy in any way. This is a standalone implementation.
This implementation is used by Dynamo to through NumPy code and lower it into PyTorch code.
To see design decisions that went into this implementation, please see the rfc.
Structure of the code
This folder exports a drop-in replacement for the NumPy namespace and its modules linalg, fft and random via its __init__.py.
The implementation is split into files that work with PyTorch objects (PyTorch Tensors, dtypes, etc) and files that
use these PyTorch-only files and convert them into functions/objects that can process all the types that the NumPy functions
accept. In particular, they accept torch._numpy.dtypes or torch._numpy.ndarrays.
The PyTorch-only files are the *_impl.py files, while the wrapper files are those that do not have an *_impl.py. This creates a
hierarchy, wherein, for example, _dtypes.py will import _dtypes_impl.py, but not the other way around. In particular, *_impl.py
will only depend on other *_impl.py files.
As discussed in the rfc, we use types as tags in our PyTorch implementations. We then use
a decorator called normalizer that will inspect these types and preprocess the inputs before sending them to the function. This
preprocessing is the one in charge of mapping array-like objects into Tensors, dtype-like objects into PyTorch dtypes, implement
the out= behaviour and so on.
In the files _funcs.py and _ufuncs.py we use register the normalizer decorator to all the *_impl.py functions.
In the file _ndarray.py we define the ndarray class, which is just a thin wrapper around a PyTorch tensor. We use the free functions
and a bit of metaprogramming to implement many of the methods.
Adding a new function
You just need to add a function in the relevant *_impl.py file. You will need to tag the inputs with the relevant Types. After that, you
can assume that the inputs are all PyTorch objects. Your function should return PyTorch tensors. The normalizer will make sure that you
always get PyTorch objects. If in doubt, you can see the implementation of the normalization attached to each type annotation in the file
_normalizations.py.
Debugging
It may be useful to figure out whether a given bug is caused by dynamo or the compatibility layer. You may use the compat layer in eager mode
simply by changing import numpy as np by import torch._numpy as np in your program, without having to call torch.compile at all.
Note that torch._numpy will be quite slow when used in eager mode, and it is in no way a replacement or an alternative to the regular PyTorch API.
This should only be used as a debugging tool.