Update Jan 13 2025:

Now that enablement in trunk has been done and 3.13/3.13t are officially released, revamping this to be a support matrix for what works today and what we plan on adding:

Source build

Nightlies (As of Jan 13 2025):

Release 2.5.0

Release 2.5.1

Release 2.6

Release 2.7 (tentative)

Original post

Python 3.13 is coming out in the near future - RC1 is planned for July 30 (first point at which the ABI is stable and packages can start releasing wheels for 3.13), and the fina release is planned for 1 Oct 2024 (see PEP 719).

I see that work on supporting 3.13 has started (gh-121979, gh-130030; Cc @atalman), but I couldn't find a tracking issue for the current status of it, so I thought it'd be useful to have one. The other motivation to open an issue now: we're working on implementing support for free-threaded CPython (a.k.a. "no-GIL", PEP 703) and are interested to start work on a first package that depends on PyTorch. Support for the default CPython 3.13 is a pre-requisite for starting that work though.

Here is a hopefully useful breakdown in steps to support Python 3.13:

• PyTorch builds from source on 3.13 (done in Changes to compile with 3.13 #126033 )
• Nightly cp313 wheels available for all platforms (waiting until final cpython rc)
• torch.compile support for 3.13 (complete @williamwen42 )
• PyTorch builds from source with free-threaded CPython (blocked on dependency support + update in core, tracking PR [Not to be merged] Update to compile and run test_torch with 3.13t #130689)
• cp313t (free-threaded) nightly wheels available (blocked on ft build and dependencies)

Python 3.13 CD:

• Nightly cp313 Linux wheels available: Add linux manywheel python 3.13 binary workflows #130030
• Windows python 3.13 update smoke_test builder#2035
• Nightly windows wheels : CD Enable Python 3.13 on windows #138095
• Nightly macos: [CD] Enable 3.13 triton build #140137
• Nightly aarch64 wheels : [Python 3.13 CD] Enable Aarch64 py3.13 builds #138629
• Add python 3.13.0t build to Docker images #136001
• Add py3.13t linux wheel build #137127
• Python 3.13t changes required to build python 3.13t wheel builder#2001
• Enable py3.13 wheels for ROCm #142294

Python 3.13 CI & torch.compile:

• Linux Wheels: Remove triton dependency python < 3.13 constraint #143162
• Python 3.11, 3.12 Remove tests covered by 3.13 #143078
• Python 3.13 CI add crossref test to existing linux-focal-py3_13-clang10-build  #143074
• [CI, 3.13] enable dynamo_wrapped unittests in 3.13 #141264
• [ci, 3.13] disable segfaulting dynamo-wrapped profiler test #141951
• [ci, 3.13] fix/skip failing numpy 2.0+ dynamo-wrapped tests #141950
• [ci, 3.13] skip failing module tracker dynamo-wrapped test #141887
• [ci, 3.13] skip failing torch.package dynamo-wrapped test #141886
• [dynamo, 3.13] disable 3.13.0 warning in dynamo-wrapped tests #141860
• [dynamo, 3.13] add JUMP_BACKWARD_NO_INTERRUPT to terminal opcodes #141859
• [ci] add 3.13 inductor unittests to CI #140733
• [CI, 3.13] enable 3.13 CI #139533
• [ci, 3.13] skip dynamo-xpass'd numpy tests in numpy >= 2.0 #141862
• [dynamo, 3.13] support CONVERT_VALUE #141858
• [ci, 3.13] disable segfaulting profiler tests in 3.13 #141674
• [ci, 3.13] disable another failing cpp_extension test in 3.13 #141673
• [ci, 3.13] disable failing cpp_extension test due to weights_only error in numpy 2.1 #141623
• [ci, 3.13] disable some quantization tests affected by numpy 2.1 overflow error #141621
• [ci, 3.13] skip some parts of a failing jit test in 3.13 #141605
• [ci, 3.13] update test_testing.py usage of locals() for 3.13 #141577
• [ci, 3.13] update tensorboard version for 3.13 to fix broken tests #141572
• format test_monitor.py and test_tensorboard.py #142003
• [torch.package, 3.13] fixes to torch.package for 3.13 #141409

torch.compile is broken out separately based on the history for Python 3.12 (see gh-110436 for general support, and gh-120233 for torch.compile).

The status of dependencies should be fine AFAIK. E.g. Pybind11 2.13.1 has full support, and SymPy 1.13.0rc4 has too (xref sympy#26692).

For free-threaded support we should probably open a separate tracking issue once regular 3.13 support is in good shape. I'll note here that Pybind11 and NumPy already have mostly-complete support for it, and NumPy should release (experimental) wheels shortly after Python 3.13.0rc1).

cc @malfet @seemethere @pytorch/pytorch-dev-infra @albanD