num_rand_blocks: int. Number of random chunks per row. batch_size: int. Batch size for computation. from_seq_length: int. length of from sequence. from_block_size: int. size of block in from sequence. Returns: float Tensor of shape [batch_size, num_attention_heads, from_seq_length//from_block_size-2, from_block_size, num_rand_blocks*to_block_size]. rO