aboutsummaryrefslogtreecommitdiff
path: root/embed_internal.h
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-15 10:40:44 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-15 10:40:44 +0200
commit6fd890a6ba42f9bd4683fc2ccd6243872b163e67 (patch)
tree035881de454f05ab86e1859364d16290e9f791e8 /embed_internal.h
parent07d3b036de430fb86935e029ca03bcf6011a02e1 (diff)
downloadffwd-6fd890a6ba42f9bd4683fc2ccd6243872b163e67.tar.gz
ffwd-6fd890a6ba42f9bd4683fc2ccd6243872b163e67.zip
late: add CUDA late-interaction rerank path (GPU encode, host MaxSim)
Diffstat (limited to 'embed_internal.h')
-rw-r--r--embed_internal.h7
1 files changed, 7 insertions, 0 deletions
diff --git a/embed_internal.h b/embed_internal.h
index d2c1a30..7d6977c 100644
--- a/embed_internal.h
+++ b/embed_internal.h
@@ -81,4 +81,11 @@ struct embed_workspace {
size_t bf16_widen_count;
};
+/* Internal accessors for the late-interaction model. GPU backends build a device
+ * context from an already-loaded CPU late model because the model dir is a late
+ * snapshot that the pooled loader rejects; these expose the base model and the
+ * projection weight without making struct embed_late_model public. */
+embed_model_t *embed_late_model_base(const embed_late_model_t *model);
+const embed_weight_ref_t *embed_late_model_projection(const embed_late_model_t *model);
+
#endif /* EMBED_INTERNAL_H */