From 6fd890a6ba42f9bd4683fc2ccd6243872b163e67 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Mon, 15 Jun 2026 10:40:44 +0200 Subject: late: add CUDA late-interaction rerank path (GPU encode, host MaxSim) --- embed.c | 10 ++++++++++ 1 file changed, 10 insertions(+) (limited to 'embed.c') diff --git a/embed.c b/embed.c index 56a9381..645660c 100644 --- a/embed.c +++ b/embed.c @@ -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, -- cgit v1.3.1