Functions§
- ground_
truth_ block_ weights - 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.
- predictor_
score_ ref - project_
kv_ block_ ref - Per-token projection. Output layout
[num_kv_heads, block_size, r]. - project_
q_ ref