[FIX] VLLM reasoning with mm: is a part of context

Minimax have custom mm: block of reasoning, so i fixed block to exclude it only {{- think_end_token }} is emitted the actual reasoning content is stripped from the prompt sent to the model.

Why: Multi-turn conversations with full reasoning rendered each turn caused context window bloat. Now the model gets a signal (think_end_token) that thinking occurred, but the reasoning text itself is omitted from history to save tokens/context space.

chat_template.jinja

@@ -159,13 +159,12 @@
             {%- endif %}
         {%- endif %}
 
-        {%- if reasoning_content -%}
-            {#- Render thinking for every assistant turn (all-turn visible) -#}
-            {{- think_begin_token ~ reasoning_content ~ think_end_token }}
-        {%- else -%}
-            {#- No thinking rendered → prefix with think_end_token -#}
-            {{- think_end_token }}
-        {%- endif -%}
+        {#- PATCH: strip reasoning from history to save context.
+            Only emit think_end_token prefix so the model knows the
+            (omitted) thinking block ended. Original behavior rendered
+            <mm:think>...</mm:think> for every assistant turn, which
+            caused context bloat in multi-turn conversations. -#}
+        {{- think_end_token }}
 
         {%- if content -%}
             {{- content }}
@@ -244,4 +243,4 @@
 {%- else -%}
     {#- adaptive: no prefix, let model decide -#}
 {%- endif -%}
-{%- endif -%}
+{%- endif -%}