Another python profile that isn't mentioned is pprofile[0]. It is also a statistical profiler with line-level granularity. It doesn't have the memory profiling abilities (that is pretty slick...), but it also doesn't require the dynamic library injection.

I don't know how it's statistical profiling speed compares to scalene, but it would be great to see the comparison.

Also, does anyone know how the malloc interacts with pytorch?

[0]https://github.com/vpelletier/pprofile