Open sleepwalker2017 opened 6 months ago
quantiles = [0.5, 0.2, 0.8] if provider == 'cublas': ms, min_ms, max_ms = triton.testing.do_bench(lambda: torch.matmul(a, b), quantiles=None)
I see the code here, it uses the median time cost instead of the avg data.
In my benchmark, the result for cublas varies a lot in different test. I don't know why.
Median is more stable than avg in statistics. It reduces outlier impact.
I see the code here, it uses the median time cost instead of the avg data.
In my benchmark, the result for cublas varies a lot in different test. I don't know why.