[Inductor] test_comprehensive_nn_functional_max_pool2d_cuda failed with Triton 3.5 #167339
Description
python test/inductor/test_torchinductor_opinfo.py -v -k test_comprehensive_nn_functional_max_pool2d_cuda_float32
Mismatched elements: 5 / 36 (13.9%)
Greatest absolute difference: 0.5730137825012207 at index (0, 0, 0, 1) (up to 1.5e-05 allowed)
Greatest relative difference: 1.3341078758239746 at index (0, 1, 0, 3) (up to 1.3e-05 allowed)
But if I force install triton 3.4 with
pip uninstall pytorch-triton -y && pip install triton==3.4
, the test will pass.
EDIT: The problem doesn't seem to appear on A100. I can reproduce on H100 and B200.
cc @chauhang @penguinwu @bertmaher @int3 @davidberard98 @nmacchioni @chenyang78 @embg @peterbell10 @aakhundov