151. Adapter Layer

Implement a bottleneck adapter module: a small residual block inserted into a frozen transformer that down-projects to a low dimension, applies a non-linearity, and up-projects back.

Signature: def adapter_forward(x: np.ndarray, W_down: np.ndarray, W_up: np.ndarray) -> np.ndarray

Shapes:

• x: (batch, d)
• W_down: (bottleneck, d)
• W_up: (d, bottleneck)

Returns: an array with the same shape as x. The block down-projects x with W_down, applies a ReLU, up-projects with W_up, and adds the residual x.