From 0a0e9f3e0fa30e49c330cc48932c703d2a4d1e7a Mon Sep 17 00:00:00 2001
From: Daniel Hiltgen <daniel@ollama.com>
Date: Tue, 19 Mar 2024 09:49:24 +0100
Subject: [PATCH] Apply 01-cache.diff

---
 llm/ext_server/server.cpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/llm/ext_server/server.cpp b/llm/ext_server/server.cpp
index 5df5bb47..f5d7863d 100644
--- a/llm/ext_server/server.cpp
+++ b/llm/ext_server/server.cpp
@@ -1007,13 +1007,15 @@ struct llama_server_context
                 slot.n_sent_text += result.text_to_send.size();
                 // add the token to slot queue and cache
             }
-            slot.add_token_string(result);
+
             if (slot.params.stream)
             {
                 send_partial_response(slot, result);
             }
         }
 
+        slot.add_token_string(result);
+
         if (incomplete)
         {
             slot.has_next_token = true;