Everyone uses pure python for purely computational tasks. numpy or pytorch has far too few operation to even count as all the computational task. e.g. most of the operations of pandas is written in pure python, and at times I found using specialised libraries could give 10x improvement but with blow to developer experience compared to python.