aboutsummaryrefslogtreecommitdiff
path: root/embed.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-15 09:45:22 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-15 09:45:22 +0200
commit43fdb308879fb0c61cb71553d9ac885c011d3a66 (patch)
treef33b036fa5ef8d8c069a5787cac684e85d54c17f /embed.h
parent16d2bd83d24d45828fe039d593f5560a0eb8329f (diff)
downloadffwd-43fdb308879fb0c61cb71553d9ac885c011d3a66.tar.gz
ffwd-43fdb308879fb0c61cb71553d9ac885c011d3a66.zip
late: batch CPU rerank candidate encoding in one packed forward
Diffstat (limited to 'embed.h')
-rw-r--r--embed.h19
1 files changed, 19 insertions, 0 deletions
diff --git a/embed.h b/embed.h
index 24e53e1..b84f71c 100644
--- a/embed.h
+++ b/embed.h
@@ -287,6 +287,25 @@ EMBED_API int embed_late_model_encode_tokens(const embed_late_model_t *model,
float *out_vectors);
/*
+ * Encode n_docs documents in one block-diagonal forward and pack each
+ * document's kept token vectors back-to-back into out_vectors[total_keep,
+ * token_dim] in document order, where total_keep is the sum of n_keep.
+ * out_offsets[n_docs + 1] receives the prefix sum of kept counts, the layout
+ * embed_late_maxsim_batch consumes. keep[d] lists the n_keep[d] token indices to
+ * retain from document d. Returns 0 on success, -1 on error.
+ */
+EMBED_API int embed_late_model_encode_docs(const embed_late_model_t *model,
+ embed_late_workspace_t *ws,
+ const int *const *doc_ids,
+ const int *n_tokens,
+ const int *const *keep,
+ const int *n_keep,
+ int n_docs,
+ int normalize,
+ float *out_vectors,
+ int *out_offsets);
+
+/*
* Sum over query tokens of max dot product against document tokens.
*/
EMBED_API float embed_late_maxsim(const float *query_vectors,