torch.js has not been released yet.

torch.BroadcastShape

export type BroadcastShape<A extends Shape, B extends Shape> =
  // If either is dynamic, result is dynamic (can't statically check)
  number extends A[number]
    ? DynamicShape
    : number extends B[number]
      ? DynamicShape
      : // Propagate error types - errors in either operand propagate to result
        A extends { readonly __isShapeError: true }
        ? A
        : B extends { readonly __isShapeError: true }
          ? B
          : // Handle empty shapes
            A extends readonly []
            ? B
            : B extends readonly []
              ? A
              : // Same length - broadcast element-wise
                A['length'] extends B['length']
                ? BroadcastSameLength<A, B>
                : // Different lengths - pad shorter one with 1s
                  A['length'] extends number
                  ? B['length'] extends number
                    ? MaxLength<A, B> extends infer L extends number
                      ? BroadcastSameLength<PadLeft<A, L>, PadLeft<B, L>>
                      : DynamicShape
                    : DynamicShape
                  : DynamicShape;

Aextends Shape

Bextends Shape

Compute broadcast shape for binary operations.

Broadcasting rules (from right to left):

Dimensions must be equal OR one of them is 1
Missing dimensions are treated as 1 (left-padded)

BroadcastShapeRule