[Bug]: "Illegal memory access was encountered" at vertical_slash_kernel

Describe the bug

Hi! Thanks for the great work! I encountered this error when trying to use vertical_slash_sparse_attention, any possible solution would be greatly appreciated!

  File "/path/MInference/minference/ops/pit_sparse_flash_attention_v2.py", line 203, in vertical_slash_sparse_attention
    out = _triton_mixed_sparse_attention(
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/path/MInference/minference/ops/pit_sparse_flash_attention_v2.py", line 153, in _triton_mixed_sparse_attention
    o = torch.zeros_like(q)
        ^^^^^^^^^^^^^^^^^^^
RuntimeError: CUDA error: an illegal memory access was encountered
Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.

Steps to reproduce

I try to create some extra pytorch tensors before using the kernel, and the more operations I make, the more easily this error is reproduced. Is it possible that some tensor operations contradict with the triton kernel?

Expected Behavior

No response

Logs

No response

Additional Information

No response