pub fn zero_padded_rows(embeddings: &Tensor, mask: &Tensor) -> Result<Tensor>Expand description
Zero every row of embeddings that sits at a padded position.
embeddings is (batch, seq_len, d), mask is (batch, seq_len) with
1 at real tokens. A padded query still produces a full output row — the
key-side mask only stops other positions attending to it — and that row
is arbitrary rather than meaningfully zero. Zeroing it makes
PlmRunner::embed_batch deterministic regardless of what each
architecture happens to compute there, so pooling over the sequence axis
does not silently mix garbage into the result.