Elementwise

elementwise.add

tensor.add — elementwise a + b, numpy-style broadcasting.

Signature: elementwise.add((a, b))

elementwise.add_backward

tensor.add_backward — da = g, db = g, each unbroadcast.

Signature: elementwise.add_backward((a, b, g))

elementwise.div

tensor.div — elementwise a / b, numpy-style broadcasting. x/0

is inf/NaN, matching the scalar / (and therefore the host op — the serde_json wire then carries null, which both sides reject on re-decode identically).

Signature: elementwise.div((a, b))

elementwise.div_backward

tensor.div_backward — da = g/b, db = −g·a/b², unbroadcast.

Signature: elementwise.div_backward((a, b, g))

elementwise.mul

tensor.mul — elementwise a * b, numpy-style broadcasting.

Signature: elementwise.mul((a, b))

elementwise.mul_backward

tensor.mul_backward — da = g·b, db = g·a, unbroadcast.

Signature: elementwise.mul_backward((a, b, g))

elementwise.sub

tensor.sub — elementwise a - b, numpy-style broadcasting.

Signature: elementwise.sub((a, b))

elementwise.sub_backward

tensor.sub_backward — da = g, db = −g, unbroadcast.

Signature: elementwise.sub_backward((a, b, g))

elementwise.unbroadcast

unbroadcast(gd, gsh, tsh) — the adjoint of broadcasting: sum the

flat grad gd (shape gsh) down to tsh, summing out leading dims tsh lacks and any dim where tsh is 1 but gsh is larger (mirrors unbroadcast in host-tensor's elementwise.rs).

Signature: elementwise.unbroadcast((gd, gsh, tsh))