Question about causal safety in DeepSeek-V4 CSA prefill retrieval
possible causality ambiguity during the prefill phase.
From the public implementation, the Lightning Indexer computes top-k retrieval over all compressed memories:
scores = q @ compressed_kv.transpose(-1, -2)
scores = F.relu(scores)
index_scores = (scores * weights.unsqueeze(-1)).sum(dim=2)
topk = index_scores.topk(k, dim=-1).indices
However, I could not find an explicit causal masking step before topk selection.
During prefill (seq_len > 1), this seems to imply that an earlier query token could potentially retrieve compressed memories corresponding to future token regions. For example:
q0 could retrieve M3
where M3 summarizes future tokens later in the prompt
I understand that later attention masking may block some retrieved entries, but I could not find discussion in the paper/blog about:
1 whether top-k retrieval itself is causally constrained
2 whether future compressed blocks are masked before retrieval
3 how causal safety is guaranteed during prefill for CSA sparse retrieval
Am I missing something or misunderstood
Would appreciate clarification from anyone familiar with the implementation details.