Function sparse_row_softmax 

pub fn sparse_row_softmax(row_offsets: &[usize], values: &[f64]) -> Vec<f64>

Row-wise softmax over sparse matrix values.

For each row defined by row_offsets, computes:

softmax(v_i) = exp(v_i - max_row) / sum_j exp(v_j - max_row)

Uses the numerically stable log-sum-exp trick.

Arguments

• row_offsets – CSR row pointer.
• values – Sparse matrix values (modified in place conceptually; returns new vector).

Returns

A new vector of softmax-normalized values with the same sparsity pattern.