From f9169667996d4753e4e32abb87dcd179f07efb08 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Wed, 11 Feb 2026 21:26:16 +0800
Subject: [PATCH] force mistral think tokens to print

---
 src/llama-vocab.cpp | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/src/llama-vocab.cpp b/src/llama-vocab.cpp
index a775db830..f9a48dd83 100644
--- a/src/llama-vocab.cpp
+++ b/src/llama-vocab.cpp
@@ -2704,6 +2704,13 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
                 attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
             }
+
+            if(t.first=="[THINK]" || t.first=="[/THINK]" || t.first=="<think>" || t.first=="</think>")
+            {
+                 LLAMA_LOG_WARN("%s: setting token '%s' (%d) attribute to USER_DEFINED (%u), old attributes: %u\n",
+                        __func__, t.first.c_str(), t.second, LLAMA_TOKEN_ATTR_USER_DEFINED, attr);
+                attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
+            }
         }
 
         // sanity checks