ML Programming Screen — QKV Attention & einsum

Requirements This interview segment may ask you to: Build a small transformer-block component from scratch in NumPy or PyTorch. The usual target is scaled dot-product attention—including QKV products, softmax, and the output projection—or a wrapper that combines several attention heads. Apply einsum correctly both when forming QKᵀ and when combining attention weights with values. Answer brief ML questions about causal and padding masks, why softmax operates along the final…