From cd5ba86fe9395cdc7527d3a57d1f852031cbdbe4 Mon Sep 17 00:00:00 2001
From: Pedro Cuenca <pedro@huggingface.co>
Date: Tue, 4 Aug 2026 23:04:06 +0200
Subject: [PATCH] Go back to using delimiters.

Otherwise our generations are worse.

Transformers does not use them. We need to trace inputs to verify
whether they are equivalent.
---
 tools/mtmd/mtmd.cpp | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/tools/mtmd/mtmd.cpp b/tools/mtmd/mtmd.cpp
index 5c49b35408..ea8f744e5f 100644
--- a/tools/mtmd/mtmd.cpp
+++ b/tools/mtmd/mtmd.cpp
@@ -472,7 +472,10 @@ struct mtmd_context {
                 } break;
             case PROJECTOR_TYPE_ONYX:
                 {
-                    // Follow transformers Onyx processing: <|patch|>*N, no delimiters
+                    // NOTE: transformers no longer uses delimiters (just <|patch|>*N),
+                    // but we get subpar generations if we don't.
+                    img_beg = "<|image_start|>";
+                    img_end = "<|image_end|>";
                     image_preproc = std::make_unique<mtmd_image_preprocessor_onyx>(ctx_v);
                 } break;
             case PROJECTOR_TYPE_YOUTUVL: