Implement Multi-Head Attention Forward Pass

Problem: Implement Multi-Head Attention Forward Pass Implement a simplified Multi-Head Self-Attention forward pass. You are given: An input tensor X with shape (B, T, D) B: batch size T: sequence length D: embedding dimension Number of heads H, where D % H == 0 Four weight matrices: Wq, shape (D, D) Wk, shape (D, D) Wv, shape (D, D) Wo, shape (D, D) Compute: Q = X @ Wq K = X @ Wk V = X @ Wv Split Q/K/V into H heads along the last dimension. Each head has dimension Dh = D / H For each batch and each head, compute scaled dot-product attention: Attention(Q, K, V) = softmax(QK^T / sqrt(Dh)) V Concatenate all heads back into shape (B, T, D) Compute the final output: O = concat_heads @ Wo You do not need to implement bias, dropout, masks, layer normalization, or backpropagation. Input Format Read from standard input: B T D H B*T*D floating-point numbers for X D*D floating-point numbers for Wq D*D floating-point numbers for Wk D*D floating-point numbers for Wv D*D floating-point numbers for Wo All matrices are provided in row-major order. Output Format Print B*T lines. Each line contains the D-dimensional output for one token. Print each floating-point number with 6 digits after the decimal point. Constraints 1 <= B <= 4 1 <= T <= 32 1 <= D <= 128 1 <= H <= D D % H == 0 Absolute value of each input float is at most 10 Example Input: 1 2 2 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 Output: 0.669762 0.330238 0.330238 0.669762 Example Input 1 1 2 1 1 2 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 Output 1.000000 2.000000