torch.js has not been released yet.

torch.MatmulShape

export type MatmulShape<A extends Shape, B extends Shape> =
  // Handle dynamic shapes
  number extends A['length']
    ? DynamicShape
    : number extends B['length']
      ? DynamicShape
      : // 1D @ 1D: dot product
        A['length'] extends 1
        ? B['length'] extends 1
          ? A[0] extends B[0]
            ? readonly []
            : matmul_error_inner_dimensions_do_not_match<A[0] & number, B[0] & number, A, B>
          : // 1D @ 2D
            B['length'] extends 2
            ? A[0] extends B[0]
              ? readonly [B[1] & number]
              : matmul_error_inner_dimensions_do_not_match<A[0] & number, B[0] & number, A, B>
            : DynamicShape
        : // 2D @ 1D
          A['length'] extends 2
          ? B['length'] extends 1
            ? A[1] extends B[0]
              ? readonly [A[0] & number]
              : matmul_error_inner_dimensions_do_not_match<A[1] & number, B[0] & number, A, B>
            : // 2D @ 2D
              B['length'] extends 2
              ? A[1] extends B[0]
                ? readonly [A[0] & number, B[1] & number]
                : matmul_error_inner_dimensions_do_not_match<A[1] & number, B[0] & number, A, B>
              : DynamicShape
          : // ND @ ND (batched) - use helper
            MatmulBatchedShape<A, B>;

Aextends Shape

Bextends Shape

Compute output shape of matmul with various input dimensions.

Handles:

2D @ 2D: [M, K] @ [K, N] → [M, N]
1D @ 1D: [K] @ [K] → [] (dot product)
2D @ 1D: [M, K] @ [K] → [M]
1D @ 2D: [K] @ [K, N] → [N]
Batched: [...B, M, K] @ [...B, K, N] → [...B, M, N] with batch broadcasting

MatmulShapeRule