pub fn ground_truth_block_weights(
q: &[bf16],
k: &[bf16],
num_q_heads: usize,
num_kv_heads: usize,
head_dim: usize,
block_size: usize,
num_blocks: usize,
) -> Vec<f32>Expand description
Ground-truth attention-weight per block: max over q_heads of the softmax-normalized attention to the block’s tokens. Used by the recall test as the “oracle” the predictor’s scores are compared against.