Fix outdated remark in shape_inference.md

lukstafi · lukstafi · commit daef2bf7c4dc · 2024-08-31T09:59:36.000+02:00
diff --git a/lib/shape_inference.md b/lib/shape_inference.md
@@ -10,7 +10,7 @@ The bulk of the projections inference happens alongside shape inference, with th
 
 A tensor shape in OCANNL is composed of three rows of axes: batch, input and output. These are ordered input-last (`batch @ output @ input`) in the underlying n-dimensional array implementation of tensors. A (fully inferred) tensor shape must have non-empty output axes; we do not use the convention where empty axes mean the tensor is a scalar -- scalars = 1-D output-only tensors. For printing and einsum-notation-like specifications, we use the syntax: `batch|input->output` (or `input->output`, `batch|output`, `output`), where `batch`, `input`, `output` are whitespace or comma or parenthesis separated axis entries; or the axis entries are the individual characters, if no separators are used (except if it's digits only).
 
-A row is a sequence of axes of a single kind: batch, input, or output. The shape type incorporates information relevant to inference, in particular shape variables: both for individual axes (`dim` variables), and for extending a row with more axes (`row` variables). Currently, all rows are (independently) broadcastable: can be broadcasted to a larger number of axes, except when used with an einsum specification that forbids it.
+A row is a sequence of axes of a single kind: batch, input, or output. The shape type incorporates information relevant to inference, in particular shape variables: both for individual axes (`dim` variables), and for extending a row with more axes (`row` variables). Currently, all rows are (independently) broadcastable: can be broadcasted to a larger number of axes. However, in OCANNL the broadcasting can happen "in the middle", with not only the given trailing axes fixed, but also with the given leading axes fixed.
 
 ```ocaml
 type dim = Var of dim_var | Dim of { d : int; label : string option; proj_id : int option }
