FusedLinearCrossEntropy got illegal memory access issue with triton 3.5.0 and PyTorch 2.9

[npuichigo]

🐛 Describe the bug

illegal memory of FusedLinearCrossEntropy

Reproduce

No response

Versions

triton 3.5.0 and PyTorch 2.9

[Tcc0403]

I couldn't reproduce the bug with current unit tests. Could you provide the code snippet?

[npuichigo]

I'm sorry that occured in my own training task and have verified that install pytorch 2.8 just works well. Additional information is it's only triggered when torch.compile is enabled.

[shimizust]

@npuichigo are you able to provide a small example that can repro?