v3.7.1-windows.post27

• Initial support of building the interpreter on Windows ARM64, see #33
• Support Windows SDK >= 10.0.28000
• More friendly error messages, see #8

Support of recent AMD consumer GPUs is generally stable. RDNA3 only has one known issue, and it will be fixed in Triton 3.8, see #7 . RDNA3.5 has passed all unit tests. RDNA4 is known to work well, although we have not run the unit test suite. You can run it and feedback any issue.

Starting from this version, the wheels are built with bitwise reproducibility and attestation. This helps prevent supply chain attacks.

v3.7.0-windows.post26

Triton 3.7 is only guaranteed to work with PyTorch 2.12 .

To install Triton 3.7 and prevent breaking with your installed PyTorch when a new version of Triton is released in future, you need to limit the version to be < 3.8:

pip install -U "triton-windows<3.8"

v3.6.0-windows.post26

• On AMD GPU, automatically find clang-cl from TheRock wheel, see #10
• Improve handling of paths longer than 260 chars by normalizing to UNC path, with the fallback of normalizing to 8.3 short names when using MSVC, see #18