From 27e76c88f0c1b33cba54dd31fade9b044719c8bc Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Sun, 31 May 2026 00:39:10 +0200 Subject: server: batch contextual document inference --- pplx_embed.h | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'pplx_embed.h') diff --git a/pplx_embed.h b/pplx_embed.h index 6a159de..3aa4059 100644 --- a/pplx_embed.h +++ b/pplx_embed.h @@ -53,6 +53,12 @@ typedef struct { int n_tokens; } pplx_span_t; +typedef struct { + pplx_input_t input; + const pplx_span_t *spans; + int n_spans; +} pplx_context_input_t; + /* ======================================================================== * API * ======================================================================== */ @@ -207,6 +213,17 @@ int pplx_model_embed_spans(const pplx_model_t *model, pplx_workspace_t *ws, const pplx_span_t *spans, int n_spans, float *out_embeddings); +/* + * Run a packed/ragged contextual document batch and pool every selected span. + * + * Documents attend independently, RoPE positions restart for each document, + * and out_embeddings contains spans in document order. + */ +int pplx_model_embed_spans_batch(const pplx_model_t *model, + pplx_workspace_t *ws, + const pplx_context_input_t *inputs, int batch, + float *out_embeddings); + /* * L2-normalize vec[dim] in place. * -- cgit v1.3.1