[BUG] ZeroDivisionError: float division by zero

[RMLYC]

Describe the Bug

In latency = do_bench( ， there is a probability that the returned latency is 0.

To Reproduce

python -m pytest tests/ops/test_mha.py

Expected Behavior

The return value latency is expected to be greater than zero.

Environment

• TileLang version: 0.1.7.post1
• PyTorch version: 2.9.0
• OS: Ubuntu22.04
• CUDA version: 12.9

[stelladuyx]

I also encountered the same problem when running the FP8 indexer test. Based on my experiments, the issue appears when the command is executed under ncu (Nsight Compute).

Observed Behavior:

1. Running the test directly without Nsight Compute works correctly.

2. The issue is unrelated to the input data. The problem still occurs after setting a fixed random seed.