Compatibility Issue: NumPy 2.0+ Causes Errors in Triton Interpret Mode

[justcyl]

Description:
Running Triton in interpret mode (TRITON_INTERPRET=1) with NumPy 2.0 or higher can lead to unexpected errors. Downgrading NumPy to a version below 2.0 (e.g., 1.21.x or 1.22.x) resolves these issues.

Steps to Fix:

1. Check your NumPy version:

   pip show numpy

2. If numpy >= 2.0, downgrade with:

   pip install numpy<2.0

[SiriusNEO]

I use numpy==2.1.2 and don't meet errors. It may also have something to do with the version of torch and triton.

Update: After a few attempts, I realized that this is indeed a compatibility issue, which may be related to the adaptation between numpy2.0 and triton.

I found two usable combinations of versions, which are:

Solution 1: Downgrade numpy to 1.26.4:

torch==2.5.0
triton==3.1.0
numpy==1.26.4

Solution 2: Use the latest nightly Triton in Github (Build it from source)

torch==2.5.0
triton==3.2.0
numpy==2.1.2

Related Issues: triton-lang/triton#4730