Skip to content
Chelis docs

Type system basics

Chelis keeps tensor dimensions and precision explicit.

Primitive ideas

Section titled “Primitive ideas”
  • No implicit precision promotion
  • No implicit broadcasting
  • Named tensor dimensions must match by name
  • Integer literals default to int32, float literals to f32

Tensor type fragments

Section titled “Tensor type fragments”
(t-tensor {} (d-name {} batch) (d-name {} seq) (t-prim {} f32))
tensor[batch, seq, f32]

A small typed program

Section titled “A small typed program”
def add_vec(x: tensor[n, f32], y: tensor[n, f32]) -> tensor[n, f32] = add(x, y)

Reading the Deep shape

Section titled “Reading the Deep shape”
(defsig {}
  add_vec
  (t-fn {}
    (t-tensor {} (d-name {} n) (t-prim {} f32))
    (t-tensor {} (d-name {} n) (t-prim {} f32))
    (t-tensor {} (d-name {} n) (t-prim {} f32))))