By the way, work group is optional but you might consider making it a required argument because the default is a poor choice. It creates a work group for each value of the first tensor which is usually very inefficient. Instead one should choose a low number of work groups and increase the number of threads per work group e.g. in GLSL: layout (local_size_x = 32) in; to make full use of the GPU's parallel architecture

Hmmm that is a good point @alexander-g. I think the default is primarily to enable less experienced users to get started with the simpler examples, but I agree that most times in produciton you'd expect people to explicitly provide these.