“Model architectures like Kimi's, which use residual connections where attention layers attend to layers several steps back, are difficult to implement using pipeline parallelism.”