Upload kimi_k2.6_fixed_template_v2.jinja

kimi_k2.6_fixed_template_v2.jinja

@@ -0,0 +1,270 @@
+{# =========================
+   Fixed Kimi K2.6 Chat Template — v2 (Generic Tool Call Format)
+   Based on original Kimi K2.6 template with improvements merged from:
+   - https://gist.github.com/fakezeta/9e8e039c60332fcb143c6e805558afe0
+   - https://github.com/allanchan339/vLLM-Qwen3-3.5-3.6-chat-template-fix
+   - https://huggingface.co/froggeric/Qwen-Fixed-Chat-Templates
+   - ubergarm's "Qwen3.6-like" approach for llama.cpp compatibility
+   
+   Fixes applied by Hunterx (https://huggingface.co/Hunterx):
+   v1 fixes (retained):
+   - Auto-close <think> before tool calls
+   - Strict tool calling instructions in system prompt
+   - Missing function name in tool call rendering
+   - Replaced .get() Python dict syntax with Jinja-native attribute access
+   - Recognizes both </think> and </thinking> as valid close tags
+   - <|think_on|> / <|think_off|> toggles
+   - developer role support
+   - LM Studio / llama.cpp Jinja runtime compatible
+
+   v2 changes:
+   - CRITICAL: Tool calls now output generic <tool_call> JSON format instead of
+     Kimi's native <|tool_call_begin|> special tokens. This allows llama.cpp,
+     ik_llama.cpp, oMLX, and LM Studio's generic tool parsers to detect and
+     route tool calls properly. The native tokens were being stripped or
+     ignored by every inference engine except vLLM.
+   - Tool call format in system prompt updated to match generic output
+   - Tool responses use generic <tool_response> tags for consistency
+   - String-form tool arguments parsed as JSON (from Qwen fix pattern)
+   ========================= #}
+
+{#- Content rendering macro (handles text, image, video) -#}
+{%- macro render_content(message) -%}
+    {%- if message.content is string -%}
+        {{ message.content }}
+    {%- elif message.content is not none -%}
+        {%- for item in message.content -%}
+            {%- if item.type == 'image' or item.type == 'image_url' -%}
+                <|media_begin|>image<|media_content|><|media_pad|><|media_end|>
+            {%- elif item.type == 'video' or item.type == 'video_url' -%}
+                <|kimi_k25_video_placeholder|>
+            {%- elif item.type == 'text' -%}
+                {{ item.text }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- Role header macro -#}
+{%- macro set_roles(message) -%}
+    {%- if message.name is defined and message.name -%}
+        {%- set role_name = message.name -%}
+    {%- else -%}
+        {%- set role_name = message.role -%}
+    {%- endif -%}
+    {%- if message.role == 'user' -%}
+        <|im_user|>{{ role_name }}<|im_middle|>
+    {%- elif message.role == 'assistant' -%}
+        <|im_assistant|>{{ role_name }}<|im_middle|>
+    {%- elif message.role == 'developer' -%}
+        <|im_system|>{{ role_name }}<|im_middle|>
+    {%- else -%}
+        <|im_system|>{{ role_name }}<|im_middle|>
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- Tool call rendering — GENERIC FORMAT for llama.cpp/oMLX/LM Studio compatibility -#}
+{%- macro render_toolcalls(message) -%}
+    {%- for tool_call in message.tool_calls -%}
+        {%- if tool_call.function is defined -%}
+            {%- set fn = tool_call.function -%}
+        {%- else -%}
+            {%- set fn = tool_call -%}
+        {%- endif -%}
+        {%- if tool_call.id is defined -%}
+            {%- set call_id = tool_call.id -%}
+        {%- else -%}
+            {%- set call_id = 'call_' ~ loop.index0 -%}
+        {%- endif -%}
+
+<tool_call>
+{"name": "{{ fn.name }}", "arguments": {% if fn.arguments is string %}{{ fn.arguments }}{% elif fn.arguments is mapping %}{{ fn.arguments | tojson }}{% elif fn.arguments is defined and fn.arguments is not none %}{{ fn.arguments | tojson }}{% else %}{}{% endif %}}
+</tool_call>
+    {%- endfor -%}
+{%- endmacro -%}
+
+{#- Thinking-toggle flag state -#}
+{%- set ns_flags = namespace(enable_thinking=true) -%}
+{%- if thinking is defined -%}
+    {%- set ns_flags.enable_thinking = thinking -%}
+{%- endif -%}
+{%- if enable_thinking is defined -%}
+    {%- set ns_flags.enable_thinking = enable_thinking -%}
+{%- endif -%}
+
+{%- set preserve_thinking = preserve_thinking | default(false) -%}
+
+{%- if not messages -%}
+    {{- raise_exception('No messages provided.') -}}
+{%- endif -%}
+
+{#- Find last non-tool-call assistant message for history/suffix split -#}
+{%- 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].tool_calls is defined -%}
+            {%- set ns.last_non_tool_call_assistant_msg = idx -%}
+            {%- break -%}
+        {%- elif messages[idx].role == 'assistant' and messages[idx].tool_calls is defined and not messages[idx].tool_calls -%}
+            {%- set ns.last_non_tool_call_assistant_msg = idx -%}
+            {%- break -%}
+        {%- endif -%}
+    {%- endfor -%}
+{%- endif -%}
+
+{#- Split messages into history (reasoning stripped) and suffix (reasoning preserved) -#}
+{%- set hist_msgs = messages[:ns.last_non_tool_call_assistant_msg + 1] -%}
+{%- set suffix_msgs = messages[ns.last_non_tool_call_assistant_msg + 1:] -%}
+
+{#- Render tool declarations with strict instructions -#}
+{%- if tools -%}
+    {%- if tools_ts_str is defined and tools_ts_str -%}
+        <|im_system|>tool_declare<|im_middle|>{{ tools_ts_str }}
+    {%- else -%}
+        <|im_system|>tool_declare<|im_middle|>{{ tools | tojson(separators=(',', ':')) }}
+    {%- endif -%}
+
+IMPORTANT: You MUST always close your thinking with </think> before emitting any tool call. Reasoning inside a tool call or its parameters is strictly forbidden and will cause a fatal error.
+
+When you need to call a tool, output one or more JSON blocks in exactly this format and nothing after the final </tool_call>:
+
+<tool_call>
+{"name": "example_function_name", "arguments": {"param1": "value1", "param2": "value2"}}
+</tool_call>
+
+Rules:
+- If a suitable tool exists for the user request, use it instead of answering from memory.
+- If the user is modifying an existing object and a prior tool response contains its id, prefer the matching update tool and reuse that id.
+- Put any reasoning or natural language BEFORE the first <tool_call>, never after the last </tool_call>.
+- Include every required parameter.
+- Preserve user-provided string values verbatim when possible for ids, names, titles, subjects, emails, and search queries.
+- For object or array parameter values, write valid JSON.
+- If no tool is needed, answer normally.<|im_end|>
+{%- endif -%}
+
+{#- ======== Render history messages (reasoning stripped) ======== -#}
+{%- for message in hist_msgs -%}
+    {%- set content = render_content(message) | trim -%}
+
+    {#- Handle think toggles -#}
+    {%- if '<|think_off|>' in content -%}
+        {%- set ns_flags.enable_thinking = false -%}
+        {%- set content = content.replace('<|think_off|>', '') -%}
+    {%- endif -%}
+    {%- if '<|think_on|>' in content -%}
+        {%- set ns_flags.enable_thinking = true -%}
+        {%- set content = content.replace('<|think_on|>', '') -%}
+    {%- endif -%}
+    {%- set content = content | trim -%}
+
+    {{ set_roles(message) }}
+    {%- if message.role == 'assistant' -%}
+        {#- Strip reasoning from history — empty think block -#}
+        <think></think>{{ content }}
+        {%- if message.tool_calls is defined and message.tool_calls -%}
+            {{ render_toolcalls(message) }}
+        {%- endif -%}
+    {%- elif message.role == 'tool' -%}
+        {%- if message.tool_call_id is defined -%}
+            {%- set tool_call_id = message.tool_call_id -%}
+        {%- else -%}
+            {%- set tool_call_id = 'unknown' -%}
+        {%- endif -%}
+<tool_response>
+{{ content }}
+</tool_response>
+    {%- elif message.role == 'system' or message.role == 'developer' -%}
+        {{ content }}
+    {%- elif message.content is not none -%}
+        {{ content }}
+    {%- endif -%}
+    <|im_end|>
+{%- endfor -%}
+
+{#- ======== Render suffix messages (reasoning preserved) ======== -#}
+{%- for message in suffix_msgs -%}
+    {%- set content = render_content(message) | trim -%}
+
+    {#- Handle think toggles -#}
+    {%- if '<|think_off|>' in content -%}
+        {%- set ns_flags.enable_thinking = false -%}
+        {%- set content = content.replace('<|think_off|>', '') -%}
+    {%- endif -%}
+    {%- if '<|think_on|>' in content -%}
+        {%- set ns_flags.enable_thinking = true -%}
+        {%- set content = content.replace('<|think_on|>', '') -%}
+    {%- endif -%}
+    {%- set content = content | trim -%}
+
+    {{ set_roles(message) }}
+    {%- if message.role == 'assistant' -%}
+        {#- Auto-close unclosed <think> before tool calls -#}
+        {%- if message.tool_calls is defined and message.tool_calls and '<think>' in content -%}
+            {%- set last_think = content.rfind('<think>') -%}
+            {%- set last_close_think = content.rfind('</think>') -%}
+            {%- set last_close_thinking = content.rfind('</thinking>') -%}
+            {%- set last_close = last_close_think if last_close_think > last_close_thinking else last_close_thinking -%}
+            {%- if last_close < last_think or last_close == -1 -%}
+                {%- set content = content + '</think>' -%}
+            {%- endif -%}
+        {%- endif -%}
+
+        {#- Extract reasoning content (recognize both </think> and </thinking>) -#}
+        {%- if ns_flags.enable_thinking is false -%}
+            <think></think>{{ content }}
+        {%- else -%}
+            {%- set reasoning_content = '' -%}
+            {%- if message.reasoning is defined and message.reasoning is string -%}
+                {%- set reasoning_content = message.reasoning -%}
+            {%- elif message.reasoning_content is defined and message.reasoning_content is string -%}
+                {%- set reasoning_content = message.reasoning_content -%}
+            {%- else -%}
+                {#- Try to extract from content -#}
+                {%- set has_think_tag = false -%}
+                {%- set think_end_token = '</think>' -%}
+                {%- if '</think>' in content -%}
+                    {%- set has_think_tag = true -%}
+                {%- elif '</thinking>' in content -%}
+                    {%- set has_think_tag = true -%}
+                    {%- set think_end_token = '</thinking>' -%}
+                {%- elif '<think>' in content -%}
+                    {%- set reasoning_content = content.split('<think>')[-1].lstrip('\n') -%}
+                    {%- set content = '' -%}
+                {%- endif -%}
+                {%- if has_think_tag -%}
+                    {%- set reasoning_content = content.split(think_end_token)[0].rstrip('\n').split('<think>')[-1].lstrip('\n') -%}
+                    {%- set content = content.split(think_end_token)[-1].lstrip('\n') -%}
+                {%- endif -%}
+            {%- endif -%}
+            <think>{{ reasoning_content }}</think>{{ content }}
+        {%- endif -%}
+        {%- if message.tool_calls is defined and message.tool_calls -%}
+            {{ render_toolcalls(message) }}
+        {%- endif -%}
+    {%- elif message.role == 'tool' -%}
+        {%- if message.tool_call_id is defined -%}
+            {%- set tool_call_id = message.tool_call_id -%}
+        {%- else -%}
+            {%- set tool_call_id = 'unknown' -%}
+        {%- endif -%}
+<tool_response>
+{{ content }}
+</tool_response>
+    {%- elif message.role == 'system' or message.role == 'developer' -%}
+        {{ content }}
+    {%- elif message.content is not none -%}
+        {{ content }}
+    {%- endif -%}
+    <|im_end|>
+{%- endfor -%}
+
+{#- ======== Generation prompt ======== -#}
+{%- if add_generation_prompt -%}
+    <|im_assistant|>assistant<|im_middle|>
+    {%- if ns_flags.enable_thinking is false -%}
+        <think></think>
+    {%- else -%}
+        <think>
+    {%- endif -%}
+{%- endif -%}
+0