force thinking & preserve_thinking in jinja; set default sampling params in generation_config.json

chat_template.jinja

@@ -14,7 +14,6 @@
       {%- endfor -%}
     {%- endif -%}
 {%- endmacro -%}
-
 {% macro set_roles(message) -%}
   {%- set role_name =  message.get('name') or  message['role'] -%}
   {%- if message['role'] == 'user' -%}
@@ -25,8 +24,6 @@
     <|im_system|>{{role_name}}<|im_middle|>
   {%- endif -%}
 {%- endmacro -%}
-
-
 {%- macro render_toolcalls(message) -%}
   <|tool_calls_section_begin|>
   {%- for tool_call in message['tool_calls'] -%}
@@ -35,24 +32,6 @@
   {%- endfor -%}
   <|tool_calls_section_end|>
 {%- endmacro -%}
-
-
-{%- set preserve_thinking = preserve_thinking | default(true) -%}
-{# Find last non-tool-call assistant message. If preserve_thinking, keep -1 so hist is empty and all msgs use suffix (retain reasoning). #}
-{%- set ns = namespace(last_non_tool_call_assistant_msg=-1) -%}
-{%- if not preserve_thinking -%}
-{%- for idx in range(messages|length-1, -1, -1) -%}
-    {%- if messages[idx]['role'] == 'assistant' and not messages[idx].get('tool_calls') -%}
-        {%- set ns.last_non_tool_call_assistant_msg = idx -%}
-        {%- break -%}
-    {%- endif -%}
-{%- endfor -%}
-{%- endif -%}
-
-{# split all messages into history & suffix, reasoning_content in suffix should be reserved.#}
-{%- set hist_msgs = messages[:ns.last_non_tool_call_assistant_msg+1] -%}
-{%- set suffix_msgs = messages[ns.last_non_tool_call_assistant_msg+1:] -%}
-
 {%- if tools -%}
   {%- if tools_ts_str -%}
     <|im_system|>tool_declare<|im_middle|>{{ tools_ts_str }}<|im_end|>
@@ -60,36 +39,13 @@
     <|im_system|>tool_declare<|im_middle|>{{ tools | tojson(separators=(',', ':')) }}<|im_end|>
   {%- endif -%}
 {%- endif -%}
-
-  
-{%- for message in hist_msgs -%}
-  {{set_roles(message)}}
-  {%- if message['role'] == 'assistant' -%}
-    <think></think>{{render_content(message)}}
-    {%- if message.get('tool_calls') -%}
-      {{render_toolcalls(message)}}
-    {%- endif -%}
-  {%- elif message['role'] == 'tool' -%}
-    {%- set tool_call_id = message.tool_call_id -%}
-    ## Return of {{ tool_call_id }}
-{{render_content(message)}}
-  {%- elif message['content'] is not none -%}
-    {{render_content(message)}}
-  {%- endif -%}
-  <|im_end|>
-{%- endfor -%}
-
-{%- for message in suffix_msgs -%}
+{%- for message in messages -%}
   {{set_roles(message)}}
   {%- if message['role'] == 'assistant' -%}
-    {%- if thinking is defined and thinking is false and preserve_thinking is false -%}
-    <think></think>{{render_content(message)}}
-    {%- else -%}
     {%- set rc = message.get('reasoning', message.get('reasoning_content', '')) -%}
     <think>{{rc}}</think>{{render_content(message)}}
-    {%- endif -%}
     {%- if message.get('tool_calls') -%}
-     {{render_toolcalls(message)}}
+      {{render_toolcalls(message)}}
     {%- endif -%}
   {%- elif message['role'] == 'tool' -%}
     {%- set tool_call_id = message.tool_call_id -%}
@@ -100,13 +56,6 @@
   {%- endif -%}
   <|im_end|>
 {%- endfor -%}
-
-
 {%- if add_generation_prompt -%}
-  <|im_assistant|>assistant<|im_middle|>
-  {%- if thinking is defined and thinking is false -%}
-  <think></think>
-  {%- else -%}
-  <think>
-  {%- endif -%}
-{%- endif -%}
+  <|im_assistant|>assistant<|im_middle|><think>
+{%- endif -%}

generation_config.json

@@ -1,4 +1,6 @@
 {
   "max_length": 262144,
-  "eos_token_id": 163586
+  "eos_token_id": 163586,
+  "temperature": 1.0,
+  "top_p": 0.95
 }

tokenization_kimi.py

@@ -328,6 +328,13 @@ class TikTokenTokenizer(PreTrainedTokenizer):
                             thinking: bool = True,
                             preserve_thinking: bool = True,
                             **kwargs):
+        """Apply chat template to conversation.
+
+        ``thinking`` and ``preserve_thinking`` are kept in the signature for
+        backward compatibility, but k27 always enables thinking and preserves
+        reasoning content. They are intentionally not forwarded to the chat
+        template.
+        """
 
         tools = deep_sort_dict(tools)
 
@@ -345,21 +352,11 @@ class TikTokenTokenizer(PreTrainedTokenizer):
         if tools_ts_str is not None:
             kwargs['tools_ts_str'] = tools_ts_str
 
-        if not thinking:
-            logger.warning("thinking=False is not supported, overriding to True")
-            thinking = True
-
-        if not preserve_thinking:
-            logger.warning("preserve_thinking=False is not supported, overriding to True")
-            preserve_thinking = True
-
         return super().apply_chat_template(
             conversation,
             tools=tools,
             tokenize=tokenize,
             add_generation_prompt=add_generation_prompt,
-            thinking=thinking,
-            preserve_thinking=preserve_thinking,
             **kwargs)