Whenever you're ready, walk me through how you'd approach this.
Implement a MHA class, together with some modeling questions related to SOTA models.
Implement multi-head scaled dot-product attention in PyTorch, including tensor reshaping, masking, dropout, and output projection. Explain the computational complexity, positional encoding choices, normalization strategy, and how you would compare Transformer variants with current state-of-the-art models. Your implementation should support batched inputs and causal or padding masks, and should be validated with shape, masking, gradient, and numerical-stability tests.