Tensor parallelism versus pipeline parallelism

A transformer is too big for one accelerator. Explain how tensor parallelism and pipeline parallelism each split it: what is cut, what crosses the wire and when, where the idle time (the pipeline bubble) comes from and how micro-batches shrink it, and how each scales with more devices. When would you use each, and how do they combine with data parallelism for training and for serving?