From 911e35bb8bb2fd1c7d3f40f27e96ff432eae7e14 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= <sigbjorn.skjaeret@scala.com>
Date: Thu, 27 Jun 2024 09:46:41 +0200
Subject: [PATCH] llama : fix CodeLlama FIM token checks (#8144)

* account for space prefix character

* use find instead
---
 src/llama.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/src/llama.cpp b/src/llama.cpp
index f78594a6f..080057332 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -5152,10 +5152,10 @@ static void llm_load_vocab(
         if (gen_name.find("code") != std::string::npos) {
             if (model.arch == LLM_ARCH_LLAMA
               && 32010 < vocab.id_to_token.size()
-              && vocab.id_to_token[32007].text == "<PRE>"
-              && vocab.id_to_token[32008].text == "<SUF>"
-              && vocab.id_to_token[32009].text == "<MID>"
-              && vocab.id_to_token[32010].text == "<EOT>") {
+              && vocab.id_to_token[32007].text.find("<PRE>") != std::string::npos
+              && vocab.id_to_token[32008].text.find("<SUF>") != std::string::npos
+              && vocab.id_to_token[32009].text.find("<MID>") != std::string::npos
+              && vocab.id_to_token[32010].text.find("<EOT>") != std::string::npos) {
                 vocab.special_prefix_id = 32007;
                 vocab.special_suffix_id = 32008;
                 vocab.special_middle_id = 32009;