einops

Four entry points: rearrange (permute, reshape, split or merge axes), reduce (the same, carrying a reduction), repeat (the same, carrying a broadcast expansion) and einsum.

Patterns are strings of axis names, so the intent is stated in the call rather than reconstructed from a run of anonymous transpose / reshape steps:

lucid.einops.rearrange(x, "b c h w -> b (h w) c") # (2,3,4,5) → (2,20,3)

The grammar is the reference einops grammar: names, parenthesised groups one level deep, ... for the axes the pattern does not name, 1 or () for a unit axis and, in reduce and repeat, a number for an unnamed axis of that length. A keyword length is also a check: an axis the input already fixes must have the length given for it, and a keyword naming no axis of the pattern is refused. A mismatch fails with a message naming the axis that did not fit, instead of producing a silently wrong shape that only surfaces layers later.

Reachable only as lucid.einops.*: there is no top-level lucid.einsum and no Tensor.rearrange method, so each op has exactly one path.