TorchFix is a Python code static analysis tool - a linter with autofix capabilities - for users of PyTorch. It can be used to find and fix issues like usage of deprecated PyTorch functions and non-public symbols, and to adopt PyTorch best practices in general.
TorchFix is built upon https://github.com/Instagram/LibCST - a library to manipulate Python concrete syntax trees. LibCST enables "codemods" (autofixes) in addition to reporting issues.
TorchFix can be used as a Flake8 plugin (linting only) or as a standalone program (with autofix available for a subset of the lint violations).
[!WARNING] Currently TorchFix is in a beta version stage, so there are still a lot of rough edges and many things can and will change.
To install the latest code from GitHub, clone/download
https://github.com/pytorch-labs/torchfix and run pip install .
inside the directory.
To install a release version from PyPI, run pip install torchfix
.
After the installation, TorchFix will be available as a Flake8 plugin, so running Flake8 normally will run the TorchFix linter.
To see only TorchFix warnings without the rest of the Flake8 linters, you can run
flake8 --isolated --select=TOR0,TOR1,TOR2
TorchFix can also be run as a standalone program: torchfix .
Add --fix
parameter to try to autofix some of the issues (the files will be overwritten!)
To see some additional debug info, add --show-stderr
parameter.
[!CAUTION] Please keep in mind that autofix is a best-effort mechanism. Given the dynamic nature of Python, and especially the beta version status of TorchFix, it's very difficult to have certainty when making changes to code, even for the seemingly trivial fixes.
Warnings for issues with codes starting with TOR0, TOR1, and TOR2 are enabled by default.
Warnings with other codes may be too noisy, so not enabled by default.
To enable them, use standard flake8 configuration options for the plugin mode or use
torchfix --select=ALL .
for the standalone mode.
If you encounter a bug or some other problem with TorchFix, please file an issue on https://github.com/pytorch-labs/torchfix/issues.
New rule codes are assigned incrementally across the following categories:
torch
functionality.pytorch/pytorch
repo, other users should not use these.TOR0, TOR1 and TOR2 are enabled by default.
This function was deprecated since PyTorch version 1.9 and is now removed.
torch.solve
is deprecated in favor of torch.linalg.solve
.
torch.linalg.solve
has its arguments reversed and does not return the LU factorization.
To get the LU factorization see torch.lu
, which can be used with torch.lu_solve
or torch.lu_unpack
.
X = torch.solve(B, A).solution
should be replaced with X = torch.linalg.solve(A, B)
.
This function was deprecated since PyTorch version 1.9 and is now removed.
torch.symeig
is deprecated in favor of torch.linalg.eigh
.
The default behavior has changed from using the upper triangular portion of the matrix by default to using the lower triangular portion.
L, _ = torch.symeig(A, upper=upper)
should be replaced with
L = torch.linalg.eigvalsh(A, UPLO='U' if upper else 'L')
and
L, V = torch.symeig(A, eigenvectors=True)
should be replaced with
L, V = torch.linalg.eigh(A, UPLO='U' if upper else 'L')
require_grad
in assignment. Did you mean requires_grad
?This is a common misspelling that can lead to silent performance issues.
use_reentrant
explicitly to checkpoint
The default value of the use_reentrant
parameter in torch.utils.checkpoint
is being changed
from True
to False
. In the meantime, the value needs to be passed explicitly.
See this forum post for details.
See TOR001
.
This function is deprecated. Use torch.nn.utils.parametrizations.weight_norm
which uses the modern parametrization API. The new weight_norm
is compatible
with state_dict
generated from old weight_norm
.
Migration guide:
The magnitude (weight_g
) and direction (weight_v
) are now expressed
as parametrizations.weight.original0
and parametrizations.weight.original1
respectively.
To remove the weight normalization reparametrization, use
torch.nn.utils.parametrize.remove_parametrizations
.
The weight is no longer recomputed once at module forward; instead, it will
be recomputed on every access. To restore the old behavior, use
torch.nn.utils.parametrize.cached
before invoking the module
in question.
This function is deprecated. Use the torch.nn.attention.sdpa_kernel
context manager instead.
Migration guide:
Each boolean input parameter (defaulting to true unless specified) of sdp_kernel
corresponds to a SDPBackened
. If the input parameter is true, the corresponding backend should be added to the input list of sdpa_kernel
.
This function is deprecated in favor of torch.linalg.multi_dot
.
Migration guide:
multi_dot
accepts a list of two or more tensors whereas chain_matmul
accepted multiple tensors as input arguments. For migration, convert the multiple tensors in argument of chain_matmul
into a list of two or more tensors for multi_dot
.
Example: Replace torch.chain_matmul(a, b, c)
with torch.linalg.multi_dot([a, b, c])
.
torch.cholesky()
is deprecated in favor of torch.linalg.cholesky()
.
Migration guide:
L = torch.cholesky(A)
should be replaced with L = torch.linalg.cholesky(A)
.L = torch.cholesky(A, upper=True)
should be replaced with L = torch.linalg.cholesky(A).mH
torch.qr()
is deprecated in favor of torch.linalg.qr()
.
Migration guide:
Q, R = torch.qr(A)
should be replaced with Q, R = torch.linalg.qr(A)
.some
of torch.qr
is replaced with a string parameter mode
in torch.linalg.qr
. The corresponding change in usage is from Q, R = torch.qr(A, some=False)
to Q, R = torch.linalg.qr(A, mode="complete")
.The function torch.range()
is deprecated as its usage is incompatible with Python's builtin range. Instead, use torch.arange()
as it produces values in [start, end)
.
Migration guide:
torch.range(start, end)
produces values in the range of [start, end]
. But torch.arange(start, end)
produces values in [start, end)
. For step size of 1, migrate usage from torch.range(start, end, 1)
to torch.arange(start, end+1, 1)
.torch.load
without weights_only
parameter is unsafe.Explicitly set weights_only
to False only if you trust the data you load and full pickle functionality is needed, otherwise set weights_only=True
.
See TOR101
.
TorchFix is BSD License licensed, as found in the LICENSE file.