diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 10:40:44 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-15 10:40:44 +0200 |
| commit | 6fd890a6ba42f9bd4683fc2ccd6243872b163e67 (patch) | |
| tree | 035881de454f05ab86e1859364d16290e9f791e8 /embed.c | |
| parent | 07d3b036de430fb86935e029ca03bcf6011a02e1 (diff) | |
| download | ffwd-6fd890a6ba42f9bd4683fc2ccd6243872b163e67.tar.gz ffwd-6fd890a6ba42f9bd4683fc2ccd6243872b163e67.zip | |
late: add CUDA late-interaction rerank path (GPU encode, host MaxSim)
Diffstat (limited to 'embed.c')
| -rw-r--r-- | embed.c | 10 |
1 files changed, 10 insertions, 0 deletions
@@ -906,6 +906,16 @@ int embed_late_model_token_dim(const embed_late_model_t *model) { return model ? model->token_dim : 0; } +/* Internal (embed_internal.h): expose the base model and projection so GPU + * backends can upload them to device without seeing struct embed_late_model. */ +embed_model_t *embed_late_model_base(const embed_late_model_t *model) { + return model ? model->base : NULL; +} + +const embed_weight_ref_t *embed_late_model_projection(const embed_late_model_t *model) { + return model ? &model->projection : NULL; +} + int embed_late_model_encode_tokens(const embed_late_model_t *model, embed_late_workspace_t *ws, const int *token_ids, |
