Tutorial 09-experimental-block-pointer with enabled block pointer fails with LTS driver (803.45)

[pbchekin]

cd python/tutorials
export TRITON_INTEL_ENABLE_BLOCK_PTR=1
python 09-experimental-block-pointer.py

triton_output=tensor([[ 0.0000e+00,  0.0000e+00,  0.0000e+00,  ...,  0.0000e+00,
          0.0000e+00,  0.0000e+00],
        [        nan, -1.7147e+38,  2.8685e-42,  ..., -4.2444e+01,
❌ Triton and Torch differ
         -5.5285e+01, -3.6599e+01],
        [        nan, -1.7147e+38,  2.8685e-42,  ...,  4.4992e+00,
         -3.3115e+01, -5.1707e+00],
        ...,
        [ 3.8032e-01, -3.9251e-07, -1.9164e-01,  ...,  1.0298e-02,
          2.1799e-01,  1.3573e-05],
        [-1.4565e-04,  6.1115e-08, -1.3391e-06,  ...,  2.8589e-02,
          5.9731e-03,  1.8974e-03],
        [-2.1044e-06, -5.5068e-09, -1.7565e-04,  ..., -2.9086e-16,
         -6.5389e-10, -1.5254e-06]], device='xpu:0')
torch_output=tensor([[ 12.5781,  14.3203,  -2.5781,  ...,   1.4473,  13.2656, -21.2656],
        [ 17.9688, -57.6562,   9.9688,  ...,  -6.9219,  20.2656,  24.0781],
        [-29.9844, -11.1094,  -9.3438,  ..., -42.4375, -55.2812, -36.5938],
        ...,
        [ -9.6406,  22.7344,   5.8633,  ...,   7.1055,  22.2656,  11.6406],
        [-24.3125,   2.1738, -17.9844,  ..., -17.2344,  15.8203,  -7.0195],
        [ 18.5156, -15.4766, -24.4062,  ...,   7.1484, -20.0156, -10.5859]],
       device='xpu:0')

Workflow run: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/9164197371/job/25194970748#step:17:232

[pbchekin]

Potentially related to the Triton cache. With Agama 881.12 the following causes segmentation fault:

rm -rf ~/.triton/cache
python 09-experimental-block-pointer.py
TRITON_INTEL_ENABLE_BLOCK_PTR=1 python 09-experimental-block-pointer.py

Update: this was a different issue fixed by #1183.

[pbchekin]

Also fails with the latest LTS driver 803.58.