spicyneuron commited on
Commit
f6423b6
·
verified ·
1 Parent(s): 8d5a7f5

Add files using upload-large-folder tool

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: mlx
3
+ license: apache-2.0
4
+ license_link: https://huggingface.co/Qwen/Qwen3-Coder-Next/blob/main/LICENSE
5
+ pipeline_tag: text-generation
6
+ base_model: Qwen/Qwen3-Coder-Next
7
+ tags:
8
+ - mlx
9
+ ---
chat_template.jinja ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {% macro render_extra_keys(json_dict, handled_keys) %}
2
+ {%- if json_dict is mapping %}
3
+ {%- for json_key in json_dict if json_key not in handled_keys %}
4
+ {%- if json_dict[json_key] is string %}
5
+ {{-'\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | string) ~ '</' ~ json_key ~ '>' }}
6
+ {%- else %}
7
+ {{- '\n<' ~ json_key ~ '>' ~ (json_dict[json_key] | tojson | safe) ~ '</' ~ json_key ~ '>' }}
8
+ {%- endif %}
9
+ {%- endfor %}
10
+ {%- endif %}
11
+ {%- endmacro %}
12
+
13
+ {%- if messages[0]["role"] == "system" %}
14
+ {%- set system_message = messages[0]["content"] %}
15
+ {%- set loop_messages = messages[1:] %}
16
+ {%- else %}
17
+ {%- set loop_messages = messages %}
18
+ {%- endif %}
19
+
20
+ {%- if not tools is defined %}
21
+ {%- set tools = [] %}
22
+ {%- endif %}
23
+
24
+ {%- if system_message is defined %}
25
+ {{- "<|im_start|>system\n" + system_message }}
26
+ {%- else %}
27
+ {%- if tools is iterable and tools | length > 0 %}
28
+ {{- "<|im_start|>system\nYou are Qwen, a helpful AI assistant that can interact with a computer to solve tasks." }}
29
+ {%- endif %}
30
+ {%- endif %}
31
+ {%- if tools is iterable and tools | length > 0 %}
32
+ {{- "\n\n# Tools\n\nYou have access to the following functions:\n\n" }}
33
+ {{- "<tools>" }}
34
+ {%- for tool in tools %}
35
+ {%- if tool.function is defined %}
36
+ {%- set tool = tool.function %}
37
+ {%- endif %}
38
+ {{- "\n<function>\n<name>" ~ tool.name ~ "</name>" }}
39
+ {%- if tool.description is defined %}
40
+ {{- '\n<description>' ~ (tool.description | trim) ~ '</description>' }}
41
+ {%- endif %}
42
+ {{- '\n<parameters>' }}
43
+ {%- if tool.parameters is defined and tool.parameters is mapping and tool.parameters.properties is defined and tool.parameters.properties is mapping %}
44
+ {%- for param_name, param_fields in tool.parameters.properties|items %}
45
+ {{- '\n<parameter>' }}
46
+ {{- '\n<name>' ~ param_name ~ '</name>' }}
47
+ {%- if param_fields.type is defined %}
48
+ {{- '\n<type>' ~ (param_fields.type | string) ~ '</type>' }}
49
+ {%- endif %}
50
+ {%- if param_fields.description is defined %}
51
+ {{- '\n<description>' ~ (param_fields.description | trim) ~ '</description>' }}
52
+ {%- endif %}
53
+ {%- set handled_keys = ['name', 'type', 'description'] %}
54
+ {{- render_extra_keys(param_fields, handled_keys) }}
55
+ {{- '\n</parameter>' }}
56
+ {%- endfor %}
57
+ {%- endif %}
58
+ {%- set handled_keys = ['type', 'properties'] %}
59
+ {{- render_extra_keys(tool.parameters, handled_keys) }}
60
+ {{- '\n</parameters>' }}
61
+ {%- set handled_keys = ['type', 'name', 'description', 'parameters'] %}
62
+ {{- render_extra_keys(tool, handled_keys) }}
63
+ {{- '\n</function>' }}
64
+ {%- endfor %}
65
+ {{- "\n</tools>" }}
66
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
67
+ {%- endif %}
68
+ {%- if system_message is defined %}
69
+ {{- '<|im_end|>\n' }}
70
+ {%- else %}
71
+ {%- if tools is iterable and tools | length > 0 %}
72
+ {{- '<|im_end|>\n' }}
73
+ {%- endif %}
74
+ {%- endif %}
75
+ {%- for message in loop_messages %}
76
+ {%- if message.role == "assistant" and message.tool_calls is defined and message.tool_calls is iterable and message.tool_calls | length > 0 %}
77
+ {{- '<|im_start|>' + message.role }}
78
+ {%- if message.content is defined and message.content is string and message.content | trim | length > 0 %}
79
+ {{- '\n' + message.content | trim + '\n' }}
80
+ {%- endif %}
81
+ {%- for tool_call in message.tool_calls %}
82
+ {%- if tool_call.function is defined %}
83
+ {%- set tool_call = tool_call.function %}
84
+ {%- endif %}
85
+ {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
86
+ {%- if tool_call.arguments is defined %}
87
+ {%- for args_name, args_value in tool_call.arguments|items %}
88
+ {{- '<parameter=' + args_name + '>\n' }}
89
+ {%- set args_value = args_value if args_value is string else args_value | tojson | safe %}
90
+ {{- args_value }}
91
+ {{- '\n</parameter>\n' }}
92
+ {%- endfor %}
93
+ {%- endif %}
94
+ {{- '</function>\n</tool_call>' }}
95
+ {%- endfor %}
96
+ {{- '<|im_end|>\n' }}
97
+ {%- elif message.role == "user" or message.role == "system" or message.role == "assistant" %}
98
+ {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
99
+ {%- elif message.role == "tool" %}
100
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
101
+ {{- '<|im_start|>user' }}
102
+ {%- endif %}
103
+ {{- '\n<tool_response>\n' }}
104
+ {{- message.content }}
105
+ {{- '\n</tool_response>' }}
106
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
107
+ {{- '<|im_end|>\n' }}
108
+ {%- elif loop.last %}
109
+ {{- '<|im_end|>\n' }}
110
+ {%- endif %}
111
+ {%- else %}
112
+ {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' }}
113
+ {%- endif %}
114
+ {%- endfor %}
115
+ {%- if add_generation_prompt %}
116
+ {{- '<|im_start|>assistant\n' }}
117
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,3077 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3NextForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0,
7
+ "bos_token_id": 151643,
8
+ "decoder_sparse_step": 1,
9
+ "eos_token_id": [
10
+ 151645,
11
+ 151643
12
+ ],
13
+ "full_attention_interval": 4,
14
+ "head_dim": 256,
15
+ "hidden_act": "silu",
16
+ "hidden_size": 2048,
17
+ "initializer_range": 0.02,
18
+ "intermediate_size": 5120,
19
+ "linear_conv_kernel_dim": 4,
20
+ "linear_key_head_dim": 128,
21
+ "linear_num_key_heads": 16,
22
+ "linear_num_value_heads": 32,
23
+ "linear_value_head_dim": 128,
24
+ "max_position_embeddings": 262144,
25
+ "mlp_only_layers": [],
26
+ "model_type": "qwen3_next",
27
+ "moe_intermediate_size": 512,
28
+ "norm_topk_prob": true,
29
+ "num_attention_heads": 16,
30
+ "num_experts": 512,
31
+ "num_experts_per_tok": 10,
32
+ "num_hidden_layers": 48,
33
+ "num_key_value_heads": 2,
34
+ "output_router_logits": false,
35
+ "partial_rotary_factor": 0.25,
36
+ "quantization": {
37
+ "group_size": 64,
38
+ "bits": 4,
39
+ "mode": "affine",
40
+ "model.embed_tokens": {
41
+ "group_size": 64,
42
+ "bits": 4,
43
+ "mode": "affine"
44
+ },
45
+ "model.layers.0.linear_attn.in_proj_qkvz": {
46
+ "group_size": 32,
47
+ "bits": 4,
48
+ "mode": "mxfp4"
49
+ },
50
+ "model.layers.0.linear_attn.in_proj_ba": {
51
+ "group_size": 64,
52
+ "bits": 4,
53
+ "mode": "affine"
54
+ },
55
+ "model.layers.0.linear_attn.out_proj": {
56
+ "group_size": 64,
57
+ "bits": 5,
58
+ "mode": "affine"
59
+ },
60
+ "model.layers.0.mlp.switch_mlp.gate_proj": {
61
+ "group_size": 32,
62
+ "bits": 4,
63
+ "mode": "mxfp4"
64
+ },
65
+ "model.layers.0.mlp.switch_mlp.up_proj": {
66
+ "group_size": 32,
67
+ "bits": 4,
68
+ "mode": "mxfp4"
69
+ },
70
+ "model.layers.0.mlp.switch_mlp.down_proj": {
71
+ "group_size": 32,
72
+ "bits": 4,
73
+ "mode": "mxfp4"
74
+ },
75
+ "model.layers.1.linear_attn.in_proj_qkvz": {
76
+ "group_size": 32,
77
+ "bits": 4,
78
+ "mode": "mxfp4"
79
+ },
80
+ "model.layers.1.linear_attn.in_proj_ba": {
81
+ "group_size": 64,
82
+ "bits": 4,
83
+ "mode": "affine"
84
+ },
85
+ "model.layers.1.linear_attn.out_proj": {
86
+ "group_size": 64,
87
+ "bits": 5,
88
+ "mode": "affine"
89
+ },
90
+ "model.layers.1.mlp.switch_mlp.gate_proj": {
91
+ "group_size": 32,
92
+ "bits": 4,
93
+ "mode": "mxfp4"
94
+ },
95
+ "model.layers.1.mlp.switch_mlp.up_proj": {
96
+ "group_size": 32,
97
+ "bits": 4,
98
+ "mode": "mxfp4"
99
+ },
100
+ "model.layers.1.mlp.switch_mlp.down_proj": {
101
+ "group_size": 64,
102
+ "bits": 6,
103
+ "mode": "affine"
104
+ },
105
+ "model.layers.2.linear_attn.in_proj_qkvz": {
106
+ "group_size": 32,
107
+ "bits": 4,
108
+ "mode": "mxfp4"
109
+ },
110
+ "model.layers.2.linear_attn.in_proj_ba": {
111
+ "group_size": 64,
112
+ "bits": 4,
113
+ "mode": "affine"
114
+ },
115
+ "model.layers.2.linear_attn.out_proj": {
116
+ "group_size": 64,
117
+ "bits": 5,
118
+ "mode": "affine"
119
+ },
120
+ "model.layers.2.mlp.switch_mlp.gate_proj": {
121
+ "group_size": 32,
122
+ "bits": 4,
123
+ "mode": "mxfp4"
124
+ },
125
+ "model.layers.2.mlp.switch_mlp.up_proj": {
126
+ "group_size": 32,
127
+ "bits": 4,
128
+ "mode": "mxfp4"
129
+ },
130
+ "model.layers.2.mlp.switch_mlp.down_proj": {
131
+ "group_size": 32,
132
+ "bits": 4,
133
+ "mode": "mxfp4"
134
+ },
135
+ "model.layers.3.self_attn.q_proj": {
136
+ "group_size": 32,
137
+ "bits": 4,
138
+ "mode": "mxfp4"
139
+ },
140
+ "model.layers.3.self_attn.k_proj": {
141
+ "group_size": 32,
142
+ "bits": 4,
143
+ "mode": "mxfp4"
144
+ },
145
+ "model.layers.3.self_attn.v_proj": {
146
+ "group_size": 32,
147
+ "bits": 4,
148
+ "mode": "mxfp4"
149
+ },
150
+ "model.layers.3.self_attn.o_proj": {
151
+ "group_size": 32,
152
+ "bits": 4,
153
+ "mode": "mxfp4"
154
+ },
155
+ "model.layers.3.mlp.switch_mlp.gate_proj": {
156
+ "group_size": 32,
157
+ "bits": 4,
158
+ "mode": "mxfp4"
159
+ },
160
+ "model.layers.3.mlp.switch_mlp.up_proj": {
161
+ "group_size": 32,
162
+ "bits": 4,
163
+ "mode": "mxfp4"
164
+ },
165
+ "model.layers.3.mlp.switch_mlp.down_proj": {
166
+ "group_size": 32,
167
+ "bits": 4,
168
+ "mode": "mxfp4"
169
+ },
170
+ "model.layers.4.linear_attn.in_proj_qkvz": {
171
+ "group_size": 32,
172
+ "bits": 4,
173
+ "mode": "mxfp4"
174
+ },
175
+ "model.layers.4.linear_attn.in_proj_ba": {
176
+ "group_size": 64,
177
+ "bits": 4,
178
+ "mode": "affine"
179
+ },
180
+ "model.layers.4.linear_attn.out_proj": {
181
+ "group_size": 64,
182
+ "bits": 5,
183
+ "mode": "affine"
184
+ },
185
+ "model.layers.4.mlp.switch_mlp.gate_proj": {
186
+ "group_size": 32,
187
+ "bits": 4,
188
+ "mode": "mxfp4"
189
+ },
190
+ "model.layers.4.mlp.switch_mlp.up_proj": {
191
+ "group_size": 32,
192
+ "bits": 4,
193
+ "mode": "mxfp4"
194
+ },
195
+ "model.layers.4.mlp.switch_mlp.down_proj": {
196
+ "group_size": 32,
197
+ "bits": 4,
198
+ "mode": "mxfp4"
199
+ },
200
+ "model.layers.5.linear_attn.in_proj_qkvz": {
201
+ "group_size": 32,
202
+ "bits": 4,
203
+ "mode": "mxfp4"
204
+ },
205
+ "model.layers.5.linear_attn.in_proj_ba": {
206
+ "group_size": 64,
207
+ "bits": 4,
208
+ "mode": "affine"
209
+ },
210
+ "model.layers.5.linear_attn.out_proj": {
211
+ "group_size": 64,
212
+ "bits": 5,
213
+ "mode": "affine"
214
+ },
215
+ "model.layers.5.mlp.switch_mlp.gate_proj": {
216
+ "group_size": 32,
217
+ "bits": 4,
218
+ "mode": "mxfp4"
219
+ },
220
+ "model.layers.5.mlp.switch_mlp.up_proj": {
221
+ "group_size": 32,
222
+ "bits": 4,
223
+ "mode": "mxfp4"
224
+ },
225
+ "model.layers.5.mlp.switch_mlp.down_proj": {
226
+ "group_size": 32,
227
+ "bits": 4,
228
+ "mode": "mxfp4"
229
+ },
230
+ "model.layers.6.linear_attn.in_proj_qkvz": {
231
+ "group_size": 32,
232
+ "bits": 4,
233
+ "mode": "mxfp4"
234
+ },
235
+ "model.layers.6.linear_attn.in_proj_ba": {
236
+ "group_size": 64,
237
+ "bits": 4,
238
+ "mode": "affine"
239
+ },
240
+ "model.layers.6.linear_attn.out_proj": {
241
+ "group_size": 64,
242
+ "bits": 5,
243
+ "mode": "affine"
244
+ },
245
+ "model.layers.6.mlp.switch_mlp.gate_proj": {
246
+ "group_size": 32,
247
+ "bits": 4,
248
+ "mode": "mxfp4"
249
+ },
250
+ "model.layers.6.mlp.switch_mlp.up_proj": {
251
+ "group_size": 32,
252
+ "bits": 4,
253
+ "mode": "mxfp4"
254
+ },
255
+ "model.layers.6.mlp.switch_mlp.down_proj": {
256
+ "group_size": 64,
257
+ "bits": 6,
258
+ "mode": "affine"
259
+ },
260
+ "model.layers.7.self_attn.q_proj": {
261
+ "group_size": 32,
262
+ "bits": 4,
263
+ "mode": "mxfp4"
264
+ },
265
+ "model.layers.7.self_attn.k_proj": {
266
+ "group_size": 32,
267
+ "bits": 4,
268
+ "mode": "mxfp4"
269
+ },
270
+ "model.layers.7.self_attn.v_proj": {
271
+ "group_size": 32,
272
+ "bits": 4,
273
+ "mode": "mxfp4"
274
+ },
275
+ "model.layers.7.self_attn.o_proj": {
276
+ "group_size": 32,
277
+ "bits": 4,
278
+ "mode": "mxfp4"
279
+ },
280
+ "model.layers.7.mlp.switch_mlp.gate_proj": {
281
+ "group_size": 32,
282
+ "bits": 4,
283
+ "mode": "mxfp4"
284
+ },
285
+ "model.layers.7.mlp.switch_mlp.up_proj": {
286
+ "group_size": 32,
287
+ "bits": 4,
288
+ "mode": "mxfp4"
289
+ },
290
+ "model.layers.7.mlp.switch_mlp.down_proj": {
291
+ "group_size": 32,
292
+ "bits": 4,
293
+ "mode": "mxfp4"
294
+ },
295
+ "model.layers.8.linear_attn.in_proj_qkvz": {
296
+ "group_size": 32,
297
+ "bits": 4,
298
+ "mode": "mxfp4"
299
+ },
300
+ "model.layers.8.linear_attn.in_proj_ba": {
301
+ "group_size": 64,
302
+ "bits": 4,
303
+ "mode": "affine"
304
+ },
305
+ "model.layers.8.linear_attn.out_proj": {
306
+ "group_size": 64,
307
+ "bits": 5,
308
+ "mode": "affine"
309
+ },
310
+ "model.layers.8.mlp.switch_mlp.gate_proj": {
311
+ "group_size": 32,
312
+ "bits": 4,
313
+ "mode": "mxfp4"
314
+ },
315
+ "model.layers.8.mlp.switch_mlp.up_proj": {
316
+ "group_size": 32,
317
+ "bits": 4,
318
+ "mode": "mxfp4"
319
+ },
320
+ "model.layers.8.mlp.switch_mlp.down_proj": {
321
+ "group_size": 32,
322
+ "bits": 4,
323
+ "mode": "mxfp4"
324
+ },
325
+ "model.layers.9.linear_attn.in_proj_qkvz": {
326
+ "group_size": 32,
327
+ "bits": 4,
328
+ "mode": "mxfp4"
329
+ },
330
+ "model.layers.9.linear_attn.in_proj_ba": {
331
+ "group_size": 64,
332
+ "bits": 4,
333
+ "mode": "affine"
334
+ },
335
+ "model.layers.9.linear_attn.out_proj": {
336
+ "group_size": 64,
337
+ "bits": 5,
338
+ "mode": "affine"
339
+ },
340
+ "model.layers.9.mlp.switch_mlp.gate_proj": {
341
+ "group_size": 32,
342
+ "bits": 4,
343
+ "mode": "mxfp4"
344
+ },
345
+ "model.layers.9.mlp.switch_mlp.up_proj": {
346
+ "group_size": 32,
347
+ "bits": 4,
348
+ "mode": "mxfp4"
349
+ },
350
+ "model.layers.9.mlp.switch_mlp.down_proj": {
351
+ "group_size": 32,
352
+ "bits": 4,
353
+ "mode": "mxfp4"
354
+ },
355
+ "model.layers.10.linear_attn.in_proj_qkvz": {
356
+ "group_size": 32,
357
+ "bits": 4,
358
+ "mode": "mxfp4"
359
+ },
360
+ "model.layers.10.linear_attn.in_proj_ba": {
361
+ "group_size": 64,
362
+ "bits": 4,
363
+ "mode": "affine"
364
+ },
365
+ "model.layers.10.linear_attn.out_proj": {
366
+ "group_size": 64,
367
+ "bits": 5,
368
+ "mode": "affine"
369
+ },
370
+ "model.layers.10.mlp.switch_mlp.gate_proj": {
371
+ "group_size": 32,
372
+ "bits": 4,
373
+ "mode": "mxfp4"
374
+ },
375
+ "model.layers.10.mlp.switch_mlp.up_proj": {
376
+ "group_size": 32,
377
+ "bits": 4,
378
+ "mode": "mxfp4"
379
+ },
380
+ "model.layers.10.mlp.switch_mlp.down_proj": {
381
+ "group_size": 64,
382
+ "bits": 6,
383
+ "mode": "affine"
384
+ },
385
+ "model.layers.11.self_attn.q_proj": {
386
+ "group_size": 32,
387
+ "bits": 4,
388
+ "mode": "mxfp4"
389
+ },
390
+ "model.layers.11.self_attn.k_proj": {
391
+ "group_size": 32,
392
+ "bits": 4,
393
+ "mode": "mxfp4"
394
+ },
395
+ "model.layers.11.self_attn.v_proj": {
396
+ "group_size": 32,
397
+ "bits": 4,
398
+ "mode": "mxfp4"
399
+ },
400
+ "model.layers.11.self_attn.o_proj": {
401
+ "group_size": 32,
402
+ "bits": 4,
403
+ "mode": "mxfp4"
404
+ },
405
+ "model.layers.11.mlp.switch_mlp.gate_proj": {
406
+ "group_size": 32,
407
+ "bits": 4,
408
+ "mode": "mxfp4"
409
+ },
410
+ "model.layers.11.mlp.switch_mlp.up_proj": {
411
+ "group_size": 32,
412
+ "bits": 4,
413
+ "mode": "mxfp4"
414
+ },
415
+ "model.layers.11.mlp.switch_mlp.down_proj": {
416
+ "group_size": 32,
417
+ "bits": 4,
418
+ "mode": "mxfp4"
419
+ },
420
+ "model.layers.12.linear_attn.in_proj_qkvz": {
421
+ "group_size": 32,
422
+ "bits": 4,
423
+ "mode": "mxfp4"
424
+ },
425
+ "model.layers.12.linear_attn.in_proj_ba": {
426
+ "group_size": 64,
427
+ "bits": 4,
428
+ "mode": "affine"
429
+ },
430
+ "model.layers.12.linear_attn.out_proj": {
431
+ "group_size": 64,
432
+ "bits": 6,
433
+ "mode": "affine"
434
+ },
435
+ "model.layers.12.mlp.switch_mlp.gate_proj": {
436
+ "group_size": 32,
437
+ "bits": 4,
438
+ "mode": "mxfp4"
439
+ },
440
+ "model.layers.12.mlp.switch_mlp.up_proj": {
441
+ "group_size": 32,
442
+ "bits": 4,
443
+ "mode": "mxfp4"
444
+ },
445
+ "model.layers.12.mlp.switch_mlp.down_proj": {
446
+ "group_size": 32,
447
+ "bits": 4,
448
+ "mode": "mxfp4"
449
+ },
450
+ "model.layers.13.linear_attn.in_proj_qkvz": {
451
+ "group_size": 32,
452
+ "bits": 4,
453
+ "mode": "mxfp4"
454
+ },
455
+ "model.layers.13.linear_attn.in_proj_ba": {
456
+ "group_size": 64,
457
+ "bits": 4,
458
+ "mode": "affine"
459
+ },
460
+ "model.layers.13.linear_attn.out_proj": {
461
+ "group_size": 64,
462
+ "bits": 5,
463
+ "mode": "affine"
464
+ },
465
+ "model.layers.13.mlp.switch_mlp.gate_proj": {
466
+ "group_size": 32,
467
+ "bits": 4,
468
+ "mode": "mxfp4"
469
+ },
470
+ "model.layers.13.mlp.switch_mlp.up_proj": {
471
+ "group_size": 32,
472
+ "bits": 4,
473
+ "mode": "mxfp4"
474
+ },
475
+ "model.layers.13.mlp.switch_mlp.down_proj": {
476
+ "group_size": 32,
477
+ "bits": 4,
478
+ "mode": "mxfp4"
479
+ },
480
+ "model.layers.14.linear_attn.in_proj_qkvz": {
481
+ "group_size": 32,
482
+ "bits": 4,
483
+ "mode": "mxfp4"
484
+ },
485
+ "model.layers.14.linear_attn.in_proj_ba": {
486
+ "group_size": 64,
487
+ "bits": 4,
488
+ "mode": "affine"
489
+ },
490
+ "model.layers.14.linear_attn.out_proj": {
491
+ "group_size": 64,
492
+ "bits": 5,
493
+ "mode": "affine"
494
+ },
495
+ "model.layers.14.mlp.switch_mlp.gate_proj": {
496
+ "group_size": 32,
497
+ "bits": 4,
498
+ "mode": "mxfp4"
499
+ },
500
+ "model.layers.14.mlp.switch_mlp.up_proj": {
501
+ "group_size": 32,
502
+ "bits": 4,
503
+ "mode": "mxfp4"
504
+ },
505
+ "model.layers.14.mlp.switch_mlp.down_proj": {
506
+ "group_size": 64,
507
+ "bits": 6,
508
+ "mode": "affine"
509
+ },
510
+ "model.layers.15.self_attn.q_proj": {
511
+ "group_size": 32,
512
+ "bits": 4,
513
+ "mode": "mxfp4"
514
+ },
515
+ "model.layers.15.self_attn.k_proj": {
516
+ "group_size": 32,
517
+ "bits": 4,
518
+ "mode": "mxfp4"
519
+ },
520
+ "model.layers.15.self_attn.v_proj": {
521
+ "group_size": 32,
522
+ "bits": 4,
523
+ "mode": "mxfp4"
524
+ },
525
+ "model.layers.15.self_attn.o_proj": {
526
+ "group_size": 32,
527
+ "bits": 4,
528
+ "mode": "mxfp4"
529
+ },
530
+ "model.layers.15.mlp.switch_mlp.gate_proj": {
531
+ "group_size": 32,
532
+ "bits": 4,
533
+ "mode": "mxfp4"
534
+ },
535
+ "model.layers.15.mlp.switch_mlp.up_proj": {
536
+ "group_size": 32,
537
+ "bits": 4,
538
+ "mode": "mxfp4"
539
+ },
540
+ "model.layers.15.mlp.switch_mlp.down_proj": {
541
+ "group_size": 32,
542
+ "bits": 4,
543
+ "mode": "mxfp4"
544
+ },
545
+ "model.layers.16.linear_attn.in_proj_qkvz": {
546
+ "group_size": 32,
547
+ "bits": 4,
548
+ "mode": "mxfp4"
549
+ },
550
+ "model.layers.16.linear_attn.in_proj_ba": {
551
+ "group_size": 64,
552
+ "bits": 4,
553
+ "mode": "affine"
554
+ },
555
+ "model.layers.16.linear_attn.out_proj": {
556
+ "group_size": 64,
557
+ "bits": 6,
558
+ "mode": "affine"
559
+ },
560
+ "model.layers.16.mlp.switch_mlp.gate_proj": {
561
+ "group_size": 32,
562
+ "bits": 4,
563
+ "mode": "mxfp4"
564
+ },
565
+ "model.layers.16.mlp.switch_mlp.up_proj": {
566
+ "group_size": 32,
567
+ "bits": 4,
568
+ "mode": "mxfp4"
569
+ },
570
+ "model.layers.16.mlp.switch_mlp.down_proj": {
571
+ "group_size": 32,
572
+ "bits": 4,
573
+ "mode": "mxfp4"
574
+ },
575
+ "model.layers.17.linear_attn.in_proj_qkvz": {
576
+ "group_size": 32,
577
+ "bits": 4,
578
+ "mode": "mxfp4"
579
+ },
580
+ "model.layers.17.linear_attn.in_proj_ba": {
581
+ "group_size": 64,
582
+ "bits": 4,
583
+ "mode": "affine"
584
+ },
585
+ "model.layers.17.linear_attn.out_proj": {
586
+ "group_size": 64,
587
+ "bits": 6,
588
+ "mode": "affine"
589
+ },
590
+ "model.layers.17.mlp.switch_mlp.gate_proj": {
591
+ "group_size": 32,
592
+ "bits": 4,
593
+ "mode": "mxfp4"
594
+ },
595
+ "model.layers.17.mlp.switch_mlp.up_proj": {
596
+ "group_size": 32,
597
+ "bits": 4,
598
+ "mode": "mxfp4"
599
+ },
600
+ "model.layers.17.mlp.switch_mlp.down_proj": {
601
+ "group_size": 32,
602
+ "bits": 4,
603
+ "mode": "mxfp4"
604
+ },
605
+ "model.layers.18.linear_attn.in_proj_qkvz": {
606
+ "group_size": 32,
607
+ "bits": 4,
608
+ "mode": "mxfp4"
609
+ },
610
+ "model.layers.18.linear_attn.in_proj_ba": {
611
+ "group_size": 64,
612
+ "bits": 4,
613
+ "mode": "affine"
614
+ },
615
+ "model.layers.18.linear_attn.out_proj": {
616
+ "group_size": 64,
617
+ "bits": 5,
618
+ "mode": "affine"
619
+ },
620
+ "model.layers.18.mlp.switch_mlp.gate_proj": {
621
+ "group_size": 32,
622
+ "bits": 4,
623
+ "mode": "mxfp4"
624
+ },
625
+ "model.layers.18.mlp.switch_mlp.up_proj": {
626
+ "group_size": 32,
627
+ "bits": 4,
628
+ "mode": "mxfp4"
629
+ },
630
+ "model.layers.18.mlp.switch_mlp.down_proj": {
631
+ "group_size": 32,
632
+ "bits": 4,
633
+ "mode": "mxfp4"
634
+ },
635
+ "model.layers.19.self_attn.q_proj": {
636
+ "group_size": 32,
637
+ "bits": 4,
638
+ "mode": "mxfp4"
639
+ },
640
+ "model.layers.19.self_attn.k_proj": {
641
+ "group_size": 32,
642
+ "bits": 4,
643
+ "mode": "mxfp4"
644
+ },
645
+ "model.layers.19.self_attn.v_proj": {
646
+ "group_size": 32,
647
+ "bits": 4,
648
+ "mode": "mxfp4"
649
+ },
650
+ "model.layers.19.self_attn.o_proj": {
651
+ "group_size": 32,
652
+ "bits": 4,
653
+ "mode": "mxfp4"
654
+ },
655
+ "model.layers.19.mlp.switch_mlp.gate_proj": {
656
+ "group_size": 32,
657
+ "bits": 4,
658
+ "mode": "mxfp4"
659
+ },
660
+ "model.layers.19.mlp.switch_mlp.up_proj": {
661
+ "group_size": 32,
662
+ "bits": 4,
663
+ "mode": "mxfp4"
664
+ },
665
+ "model.layers.19.mlp.switch_mlp.down_proj": {
666
+ "group_size": 32,
667
+ "bits": 4,
668
+ "mode": "mxfp4"
669
+ },
670
+ "model.layers.20.linear_attn.in_proj_qkvz": {
671
+ "group_size": 32,
672
+ "bits": 4,
673
+ "mode": "mxfp4"
674
+ },
675
+ "model.layers.20.linear_attn.in_proj_ba": {
676
+ "group_size": 64,
677
+ "bits": 4,
678
+ "mode": "affine"
679
+ },
680
+ "model.layers.20.linear_attn.out_proj": {
681
+ "group_size": 64,
682
+ "bits": 5,
683
+ "mode": "affine"
684
+ },
685
+ "model.layers.20.mlp.switch_mlp.gate_proj": {
686
+ "group_size": 32,
687
+ "bits": 4,
688
+ "mode": "mxfp4"
689
+ },
690
+ "model.layers.20.mlp.switch_mlp.up_proj": {
691
+ "group_size": 32,
692
+ "bits": 4,
693
+ "mode": "mxfp4"
694
+ },
695
+ "model.layers.20.mlp.switch_mlp.down_proj": {
696
+ "group_size": 32,
697
+ "bits": 4,
698
+ "mode": "mxfp4"
699
+ },
700
+ "model.layers.21.linear_attn.in_proj_qkvz": {
701
+ "group_size": 32,
702
+ "bits": 4,
703
+ "mode": "mxfp4"
704
+ },
705
+ "model.layers.21.linear_attn.in_proj_ba": {
706
+ "group_size": 64,
707
+ "bits": 4,
708
+ "mode": "affine"
709
+ },
710
+ "model.layers.21.linear_attn.out_proj": {
711
+ "group_size": 64,
712
+ "bits": 5,
713
+ "mode": "affine"
714
+ },
715
+ "model.layers.21.mlp.switch_mlp.gate_proj": {
716
+ "group_size": 32,
717
+ "bits": 4,
718
+ "mode": "mxfp4"
719
+ },
720
+ "model.layers.21.mlp.switch_mlp.up_proj": {
721
+ "group_size": 32,
722
+ "bits": 4,
723
+ "mode": "mxfp4"
724
+ },
725
+ "model.layers.21.mlp.switch_mlp.down_proj": {
726
+ "group_size": 32,
727
+ "bits": 4,
728
+ "mode": "mxfp4"
729
+ },
730
+ "model.layers.22.linear_attn.in_proj_qkvz": {
731
+ "group_size": 32,
732
+ "bits": 4,
733
+ "mode": "mxfp4"
734
+ },
735
+ "model.layers.22.linear_attn.in_proj_ba": {
736
+ "group_size": 64,
737
+ "bits": 4,
738
+ "mode": "affine"
739
+ },
740
+ "model.layers.22.linear_attn.out_proj": {
741
+ "group_size": 64,
742
+ "bits": 5,
743
+ "mode": "affine"
744
+ },
745
+ "model.layers.22.mlp.switch_mlp.gate_proj": {
746
+ "group_size": 32,
747
+ "bits": 4,
748
+ "mode": "mxfp4"
749
+ },
750
+ "model.layers.22.mlp.switch_mlp.up_proj": {
751
+ "group_size": 32,
752
+ "bits": 4,
753
+ "mode": "mxfp4"
754
+ },
755
+ "model.layers.22.mlp.switch_mlp.down_proj": {
756
+ "group_size": 64,
757
+ "bits": 6,
758
+ "mode": "affine"
759
+ },
760
+ "model.layers.23.self_attn.q_proj": {
761
+ "group_size": 32,
762
+ "bits": 4,
763
+ "mode": "mxfp4"
764
+ },
765
+ "model.layers.23.self_attn.k_proj": {
766
+ "group_size": 32,
767
+ "bits": 4,
768
+ "mode": "mxfp4"
769
+ },
770
+ "model.layers.23.self_attn.v_proj": {
771
+ "group_size": 32,
772
+ "bits": 4,
773
+ "mode": "mxfp4"
774
+ },
775
+ "model.layers.23.self_attn.o_proj": {
776
+ "group_size": 32,
777
+ "bits": 4,
778
+ "mode": "mxfp4"
779
+ },
780
+ "model.layers.23.mlp.switch_mlp.gate_proj": {
781
+ "group_size": 32,
782
+ "bits": 4,
783
+ "mode": "mxfp4"
784
+ },
785
+ "model.layers.23.mlp.switch_mlp.up_proj": {
786
+ "group_size": 32,
787
+ "bits": 4,
788
+ "mode": "mxfp4"
789
+ },
790
+ "model.layers.23.mlp.switch_mlp.down_proj": {
791
+ "group_size": 32,
792
+ "bits": 4,
793
+ "mode": "mxfp4"
794
+ },
795
+ "model.layers.24.linear_attn.in_proj_qkvz": {
796
+ "group_size": 32,
797
+ "bits": 4,
798
+ "mode": "mxfp4"
799
+ },
800
+ "model.layers.24.linear_attn.in_proj_ba": {
801
+ "group_size": 64,
802
+ "bits": 4,
803
+ "mode": "affine"
804
+ },
805
+ "model.layers.24.linear_attn.out_proj": {
806
+ "group_size": 64,
807
+ "bits": 5,
808
+ "mode": "affine"
809
+ },
810
+ "model.layers.24.mlp.switch_mlp.gate_proj": {
811
+ "group_size": 32,
812
+ "bits": 4,
813
+ "mode": "mxfp4"
814
+ },
815
+ "model.layers.24.mlp.switch_mlp.up_proj": {
816
+ "group_size": 32,
817
+ "bits": 4,
818
+ "mode": "mxfp4"
819
+ },
820
+ "model.layers.24.mlp.switch_mlp.down_proj": {
821
+ "group_size": 32,
822
+ "bits": 4,
823
+ "mode": "mxfp4"
824
+ },
825
+ "model.layers.25.linear_attn.in_proj_qkvz": {
826
+ "group_size": 32,
827
+ "bits": 4,
828
+ "mode": "mxfp4"
829
+ },
830
+ "model.layers.25.linear_attn.in_proj_ba": {
831
+ "group_size": 64,
832
+ "bits": 4,
833
+ "mode": "affine"
834
+ },
835
+ "model.layers.25.linear_attn.out_proj": {
836
+ "group_size": 64,
837
+ "bits": 5,
838
+ "mode": "affine"
839
+ },
840
+ "model.layers.25.mlp.switch_mlp.gate_proj": {
841
+ "group_size": 32,
842
+ "bits": 4,
843
+ "mode": "mxfp4"
844
+ },
845
+ "model.layers.25.mlp.switch_mlp.up_proj": {
846
+ "group_size": 32,
847
+ "bits": 4,
848
+ "mode": "mxfp4"
849
+ },
850
+ "model.layers.25.mlp.switch_mlp.down_proj": {
851
+ "group_size": 32,
852
+ "bits": 4,
853
+ "mode": "mxfp4"
854
+ },
855
+ "model.layers.26.linear_attn.in_proj_qkvz": {
856
+ "group_size": 32,
857
+ "bits": 4,
858
+ "mode": "mxfp4"
859
+ },
860
+ "model.layers.26.linear_attn.in_proj_ba": {
861
+ "group_size": 64,
862
+ "bits": 4,
863
+ "mode": "affine"
864
+ },
865
+ "model.layers.26.linear_attn.out_proj": {
866
+ "group_size": 64,
867
+ "bits": 5,
868
+ "mode": "affine"
869
+ },
870
+ "model.layers.26.mlp.switch_mlp.gate_proj": {
871
+ "group_size": 32,
872
+ "bits": 4,
873
+ "mode": "mxfp4"
874
+ },
875
+ "model.layers.26.mlp.switch_mlp.up_proj": {
876
+ "group_size": 32,
877
+ "bits": 4,
878
+ "mode": "mxfp4"
879
+ },
880
+ "model.layers.26.mlp.switch_mlp.down_proj": {
881
+ "group_size": 32,
882
+ "bits": 4,
883
+ "mode": "mxfp4"
884
+ },
885
+ "model.layers.27.self_attn.q_proj": {
886
+ "group_size": 32,
887
+ "bits": 4,
888
+ "mode": "mxfp4"
889
+ },
890
+ "model.layers.27.self_attn.k_proj": {
891
+ "group_size": 32,
892
+ "bits": 4,
893
+ "mode": "mxfp4"
894
+ },
895
+ "model.layers.27.self_attn.v_proj": {
896
+ "group_size": 32,
897
+ "bits": 4,
898
+ "mode": "mxfp4"
899
+ },
900
+ "model.layers.27.self_attn.o_proj": {
901
+ "group_size": 32,
902
+ "bits": 4,
903
+ "mode": "mxfp4"
904
+ },
905
+ "model.layers.27.mlp.switch_mlp.gate_proj": {
906
+ "group_size": 32,
907
+ "bits": 4,
908
+ "mode": "mxfp4"
909
+ },
910
+ "model.layers.27.mlp.switch_mlp.up_proj": {
911
+ "group_size": 32,
912
+ "bits": 4,
913
+ "mode": "mxfp4"
914
+ },
915
+ "model.layers.27.mlp.switch_mlp.down_proj": {
916
+ "group_size": 32,
917
+ "bits": 4,
918
+ "mode": "mxfp4"
919
+ },
920
+ "model.layers.28.linear_attn.in_proj_qkvz": {
921
+ "group_size": 32,
922
+ "bits": 4,
923
+ "mode": "mxfp4"
924
+ },
925
+ "model.layers.28.linear_attn.in_proj_ba": {
926
+ "group_size": 64,
927
+ "bits": 4,
928
+ "mode": "affine"
929
+ },
930
+ "model.layers.28.linear_attn.out_proj": {
931
+ "group_size": 64,
932
+ "bits": 5,
933
+ "mode": "affine"
934
+ },
935
+ "model.layers.28.mlp.switch_mlp.gate_proj": {
936
+ "group_size": 32,
937
+ "bits": 4,
938
+ "mode": "mxfp4"
939
+ },
940
+ "model.layers.28.mlp.switch_mlp.up_proj": {
941
+ "group_size": 32,
942
+ "bits": 4,
943
+ "mode": "mxfp4"
944
+ },
945
+ "model.layers.28.mlp.switch_mlp.down_proj": {
946
+ "group_size": 32,
947
+ "bits": 4,
948
+ "mode": "mxfp4"
949
+ },
950
+ "model.layers.29.linear_attn.in_proj_qkvz": {
951
+ "group_size": 32,
952
+ "bits": 4,
953
+ "mode": "mxfp4"
954
+ },
955
+ "model.layers.29.linear_attn.in_proj_ba": {
956
+ "group_size": 64,
957
+ "bits": 4,
958
+ "mode": "affine"
959
+ },
960
+ "model.layers.29.linear_attn.out_proj": {
961
+ "group_size": 64,
962
+ "bits": 5,
963
+ "mode": "affine"
964
+ },
965
+ "model.layers.29.mlp.switch_mlp.gate_proj": {
966
+ "group_size": 32,
967
+ "bits": 4,
968
+ "mode": "mxfp4"
969
+ },
970
+ "model.layers.29.mlp.switch_mlp.up_proj": {
971
+ "group_size": 32,
972
+ "bits": 4,
973
+ "mode": "mxfp4"
974
+ },
975
+ "model.layers.29.mlp.switch_mlp.down_proj": {
976
+ "group_size": 32,
977
+ "bits": 4,
978
+ "mode": "mxfp4"
979
+ },
980
+ "model.layers.30.linear_attn.in_proj_qkvz": {
981
+ "group_size": 32,
982
+ "bits": 4,
983
+ "mode": "mxfp4"
984
+ },
985
+ "model.layers.30.linear_attn.in_proj_ba": {
986
+ "group_size": 64,
987
+ "bits": 4,
988
+ "mode": "affine"
989
+ },
990
+ "model.layers.30.linear_attn.out_proj": {
991
+ "group_size": 64,
992
+ "bits": 6,
993
+ "mode": "affine"
994
+ },
995
+ "model.layers.30.mlp.switch_mlp.gate_proj": {
996
+ "group_size": 32,
997
+ "bits": 4,
998
+ "mode": "mxfp4"
999
+ },
1000
+ "model.layers.30.mlp.switch_mlp.up_proj": {
1001
+ "group_size": 32,
1002
+ "bits": 4,
1003
+ "mode": "mxfp4"
1004
+ },
1005
+ "model.layers.30.mlp.switch_mlp.down_proj": {
1006
+ "group_size": 32,
1007
+ "bits": 4,
1008
+ "mode": "mxfp4"
1009
+ },
1010
+ "model.layers.31.self_attn.q_proj": {
1011
+ "group_size": 32,
1012
+ "bits": 4,
1013
+ "mode": "mxfp4"
1014
+ },
1015
+ "model.layers.31.self_attn.k_proj": {
1016
+ "group_size": 32,
1017
+ "bits": 4,
1018
+ "mode": "mxfp4"
1019
+ },
1020
+ "model.layers.31.self_attn.v_proj": {
1021
+ "group_size": 32,
1022
+ "bits": 4,
1023
+ "mode": "mxfp4"
1024
+ },
1025
+ "model.layers.31.self_attn.o_proj": {
1026
+ "group_size": 32,
1027
+ "bits": 4,
1028
+ "mode": "mxfp4"
1029
+ },
1030
+ "model.layers.31.mlp.switch_mlp.gate_proj": {
1031
+ "group_size": 32,
1032
+ "bits": 4,
1033
+ "mode": "mxfp4"
1034
+ },
1035
+ "model.layers.31.mlp.switch_mlp.up_proj": {
1036
+ "group_size": 32,
1037
+ "bits": 4,
1038
+ "mode": "mxfp4"
1039
+ },
1040
+ "model.layers.31.mlp.switch_mlp.down_proj": {
1041
+ "group_size": 64,
1042
+ "bits": 6,
1043
+ "mode": "affine"
1044
+ },
1045
+ "model.layers.32.linear_attn.in_proj_qkvz": {
1046
+ "group_size": 32,
1047
+ "bits": 4,
1048
+ "mode": "mxfp4"
1049
+ },
1050
+ "model.layers.32.linear_attn.in_proj_ba": {
1051
+ "group_size": 64,
1052
+ "bits": 4,
1053
+ "mode": "affine"
1054
+ },
1055
+ "model.layers.32.linear_attn.out_proj": {
1056
+ "group_size": 64,
1057
+ "bits": 6,
1058
+ "mode": "affine"
1059
+ },
1060
+ "model.layers.32.mlp.switch_mlp.gate_proj": {
1061
+ "group_size": 32,
1062
+ "bits": 4,
1063
+ "mode": "mxfp4"
1064
+ },
1065
+ "model.layers.32.mlp.switch_mlp.up_proj": {
1066
+ "group_size": 32,
1067
+ "bits": 4,
1068
+ "mode": "mxfp4"
1069
+ },
1070
+ "model.layers.32.mlp.switch_mlp.down_proj": {
1071
+ "group_size": 32,
1072
+ "bits": 4,
1073
+ "mode": "mxfp4"
1074
+ },
1075
+ "model.layers.33.linear_attn.in_proj_qkvz": {
1076
+ "group_size": 32,
1077
+ "bits": 4,
1078
+ "mode": "mxfp4"
1079
+ },
1080
+ "model.layers.33.linear_attn.in_proj_ba": {
1081
+ "group_size": 64,
1082
+ "bits": 4,
1083
+ "mode": "affine"
1084
+ },
1085
+ "model.layers.33.linear_attn.out_proj": {
1086
+ "group_size": 64,
1087
+ "bits": 5,
1088
+ "mode": "affine"
1089
+ },
1090
+ "model.layers.33.mlp.switch_mlp.gate_proj": {
1091
+ "group_size": 32,
1092
+ "bits": 4,
1093
+ "mode": "mxfp4"
1094
+ },
1095
+ "model.layers.33.mlp.switch_mlp.up_proj": {
1096
+ "group_size": 32,
1097
+ "bits": 4,
1098
+ "mode": "mxfp4"
1099
+ },
1100
+ "model.layers.33.mlp.switch_mlp.down_proj": {
1101
+ "group_size": 32,
1102
+ "bits": 4,
1103
+ "mode": "mxfp4"
1104
+ },
1105
+ "model.layers.34.linear_attn.in_proj_qkvz": {
1106
+ "group_size": 32,
1107
+ "bits": 4,
1108
+ "mode": "mxfp4"
1109
+ },
1110
+ "model.layers.34.linear_attn.in_proj_ba": {
1111
+ "group_size": 64,
1112
+ "bits": 4,
1113
+ "mode": "affine"
1114
+ },
1115
+ "model.layers.34.linear_attn.out_proj": {
1116
+ "group_size": 64,
1117
+ "bits": 6,
1118
+ "mode": "affine"
1119
+ },
1120
+ "model.layers.34.mlp.switch_mlp.gate_proj": {
1121
+ "group_size": 32,
1122
+ "bits": 4,
1123
+ "mode": "mxfp4"
1124
+ },
1125
+ "model.layers.34.mlp.switch_mlp.up_proj": {
1126
+ "group_size": 32,
1127
+ "bits": 4,
1128
+ "mode": "mxfp4"
1129
+ },
1130
+ "model.layers.34.mlp.switch_mlp.down_proj": {
1131
+ "group_size": 32,
1132
+ "bits": 4,
1133
+ "mode": "mxfp4"
1134
+ },
1135
+ "model.layers.35.self_attn.q_proj": {
1136
+ "group_size": 32,
1137
+ "bits": 4,
1138
+ "mode": "mxfp4"
1139
+ },
1140
+ "model.layers.35.self_attn.k_proj": {
1141
+ "group_size": 32,
1142
+ "bits": 4,
1143
+ "mode": "mxfp4"
1144
+ },
1145
+ "model.layers.35.self_attn.v_proj": {
1146
+ "group_size": 32,
1147
+ "bits": 4,
1148
+ "mode": "mxfp4"
1149
+ },
1150
+ "model.layers.35.self_attn.o_proj": {
1151
+ "group_size": 32,
1152
+ "bits": 4,
1153
+ "mode": "mxfp4"
1154
+ },
1155
+ "model.layers.35.mlp.switch_mlp.gate_proj": {
1156
+ "group_size": 32,
1157
+ "bits": 4,
1158
+ "mode": "mxfp4"
1159
+ },
1160
+ "model.layers.35.mlp.switch_mlp.up_proj": {
1161
+ "group_size": 32,
1162
+ "bits": 4,
1163
+ "mode": "mxfp4"
1164
+ },
1165
+ "model.layers.35.mlp.switch_mlp.down_proj": {
1166
+ "group_size": 32,
1167
+ "bits": 4,
1168
+ "mode": "mxfp4"
1169
+ },
1170
+ "model.layers.36.linear_attn.in_proj_qkvz": {
1171
+ "group_size": 32,
1172
+ "bits": 4,
1173
+ "mode": "mxfp4"
1174
+ },
1175
+ "model.layers.36.linear_attn.in_proj_ba": {
1176
+ "group_size": 64,
1177
+ "bits": 4,
1178
+ "mode": "affine"
1179
+ },
1180
+ "model.layers.36.linear_attn.out_proj": {
1181
+ "group_size": 64,
1182
+ "bits": 5,
1183
+ "mode": "affine"
1184
+ },
1185
+ "model.layers.36.mlp.switch_mlp.gate_proj": {
1186
+ "group_size": 32,
1187
+ "bits": 4,
1188
+ "mode": "mxfp4"
1189
+ },
1190
+ "model.layers.36.mlp.switch_mlp.up_proj": {
1191
+ "group_size": 32,
1192
+ "bits": 4,
1193
+ "mode": "mxfp4"
1194
+ },
1195
+ "model.layers.36.mlp.switch_mlp.down_proj": {
1196
+ "group_size": 32,
1197
+ "bits": 4,
1198
+ "mode": "mxfp4"
1199
+ },
1200
+ "model.layers.37.linear_attn.in_proj_qkvz": {
1201
+ "group_size": 32,
1202
+ "bits": 4,
1203
+ "mode": "mxfp4"
1204
+ },
1205
+ "model.layers.37.linear_attn.in_proj_ba": {
1206
+ "group_size": 64,
1207
+ "bits": 4,
1208
+ "mode": "affine"
1209
+ },
1210
+ "model.layers.37.linear_attn.out_proj": {
1211
+ "group_size": 64,
1212
+ "bits": 5,
1213
+ "mode": "affine"
1214
+ },
1215
+ "model.layers.37.mlp.switch_mlp.gate_proj": {
1216
+ "group_size": 32,
1217
+ "bits": 4,
1218
+ "mode": "mxfp4"
1219
+ },
1220
+ "model.layers.37.mlp.switch_mlp.up_proj": {
1221
+ "group_size": 32,
1222
+ "bits": 4,
1223
+ "mode": "mxfp4"
1224
+ },
1225
+ "model.layers.37.mlp.switch_mlp.down_proj": {
1226
+ "group_size": 32,
1227
+ "bits": 4,
1228
+ "mode": "mxfp4"
1229
+ },
1230
+ "model.layers.38.linear_attn.in_proj_qkvz": {
1231
+ "group_size": 32,
1232
+ "bits": 4,
1233
+ "mode": "mxfp4"
1234
+ },
1235
+ "model.layers.38.linear_attn.in_proj_ba": {
1236
+ "group_size": 64,
1237
+ "bits": 4,
1238
+ "mode": "affine"
1239
+ },
1240
+ "model.layers.38.linear_attn.out_proj": {
1241
+ "group_size": 64,
1242
+ "bits": 5,
1243
+ "mode": "affine"
1244
+ },
1245
+ "model.layers.38.mlp.switch_mlp.gate_proj": {
1246
+ "group_size": 32,
1247
+ "bits": 4,
1248
+ "mode": "mxfp4"
1249
+ },
1250
+ "model.layers.38.mlp.switch_mlp.up_proj": {
1251
+ "group_size": 32,
1252
+ "bits": 4,
1253
+ "mode": "mxfp4"
1254
+ },
1255
+ "model.layers.38.mlp.switch_mlp.down_proj": {
1256
+ "group_size": 32,
1257
+ "bits": 4,
1258
+ "mode": "mxfp4"
1259
+ },
1260
+ "model.layers.39.self_attn.q_proj": {
1261
+ "group_size": 32,
1262
+ "bits": 4,
1263
+ "mode": "mxfp4"
1264
+ },
1265
+ "model.layers.39.self_attn.k_proj": {
1266
+ "group_size": 32,
1267
+ "bits": 4,
1268
+ "mode": "mxfp4"
1269
+ },
1270
+ "model.layers.39.self_attn.v_proj": {
1271
+ "group_size": 32,
1272
+ "bits": 4,
1273
+ "mode": "mxfp4"
1274
+ },
1275
+ "model.layers.39.self_attn.o_proj": {
1276
+ "group_size": 32,
1277
+ "bits": 4,
1278
+ "mode": "mxfp4"
1279
+ },
1280
+ "model.layers.39.mlp.switch_mlp.gate_proj": {
1281
+ "group_size": 32,
1282
+ "bits": 4,
1283
+ "mode": "mxfp4"
1284
+ },
1285
+ "model.layers.39.mlp.switch_mlp.up_proj": {
1286
+ "group_size": 32,
1287
+ "bits": 4,
1288
+ "mode": "mxfp4"
1289
+ },
1290
+ "model.layers.39.mlp.switch_mlp.down_proj": {
1291
+ "group_size": 64,
1292
+ "bits": 6,
1293
+ "mode": "affine"
1294
+ },
1295
+ "model.layers.40.linear_attn.in_proj_qkvz": {
1296
+ "group_size": 32,
1297
+ "bits": 4,
1298
+ "mode": "mxfp4"
1299
+ },
1300
+ "model.layers.40.linear_attn.in_proj_ba": {
1301
+ "group_size": 64,
1302
+ "bits": 4,
1303
+ "mode": "affine"
1304
+ },
1305
+ "model.layers.40.linear_attn.out_proj": {
1306
+ "group_size": 64,
1307
+ "bits": 5,
1308
+ "mode": "affine"
1309
+ },
1310
+ "model.layers.40.mlp.switch_mlp.gate_proj": {
1311
+ "group_size": 32,
1312
+ "bits": 4,
1313
+ "mode": "mxfp4"
1314
+ },
1315
+ "model.layers.40.mlp.switch_mlp.up_proj": {
1316
+ "group_size": 32,
1317
+ "bits": 4,
1318
+ "mode": "mxfp4"
1319
+ },
1320
+ "model.layers.40.mlp.switch_mlp.down_proj": {
1321
+ "group_size": 32,
1322
+ "bits": 4,
1323
+ "mode": "mxfp4"
1324
+ },
1325
+ "model.layers.41.linear_attn.in_proj_qkvz": {
1326
+ "group_size": 32,
1327
+ "bits": 4,
1328
+ "mode": "mxfp4"
1329
+ },
1330
+ "model.layers.41.linear_attn.in_proj_ba": {
1331
+ "group_size": 64,
1332
+ "bits": 4,
1333
+ "mode": "affine"
1334
+ },
1335
+ "model.layers.41.linear_attn.out_proj": {
1336
+ "group_size": 64,
1337
+ "bits": 5,
1338
+ "mode": "affine"
1339
+ },
1340
+ "model.layers.41.mlp.switch_mlp.gate_proj": {
1341
+ "group_size": 32,
1342
+ "bits": 4,
1343
+ "mode": "mxfp4"
1344
+ },
1345
+ "model.layers.41.mlp.switch_mlp.up_proj": {
1346
+ "group_size": 32,
1347
+ "bits": 4,
1348
+ "mode": "mxfp4"
1349
+ },
1350
+ "model.layers.41.mlp.switch_mlp.down_proj": {
1351
+ "group_size": 32,
1352
+ "bits": 4,
1353
+ "mode": "mxfp4"
1354
+ },
1355
+ "model.layers.42.linear_attn.in_proj_qkvz": {
1356
+ "group_size": 32,
1357
+ "bits": 4,
1358
+ "mode": "mxfp4"
1359
+ },
1360
+ "model.layers.42.linear_attn.in_proj_ba": {
1361
+ "group_size": 64,
1362
+ "bits": 4,
1363
+ "mode": "affine"
1364
+ },
1365
+ "model.layers.42.linear_attn.out_proj": {
1366
+ "group_size": 64,
1367
+ "bits": 5,
1368
+ "mode": "affine"
1369
+ },
1370
+ "model.layers.42.mlp.switch_mlp.gate_proj": {
1371
+ "group_size": 32,
1372
+ "bits": 4,
1373
+ "mode": "mxfp4"
1374
+ },
1375
+ "model.layers.42.mlp.switch_mlp.up_proj": {
1376
+ "group_size": 32,
1377
+ "bits": 4,
1378
+ "mode": "mxfp4"
1379
+ },
1380
+ "model.layers.42.mlp.switch_mlp.down_proj": {
1381
+ "group_size": 64,
1382
+ "bits": 6,
1383
+ "mode": "affine"
1384
+ },
1385
+ "model.layers.43.self_attn.q_proj": {
1386
+ "group_size": 64,
1387
+ "bits": 5,
1388
+ "mode": "affine"
1389
+ },
1390
+ "model.layers.43.self_attn.k_proj": {
1391
+ "group_size": 64,
1392
+ "bits": 5,
1393
+ "mode": "affine"
1394
+ },
1395
+ "model.layers.43.self_attn.v_proj": {
1396
+ "group_size": 64,
1397
+ "bits": 5,
1398
+ "mode": "affine"
1399
+ },
1400
+ "model.layers.43.self_attn.o_proj": {
1401
+ "group_size": 64,
1402
+ "bits": 5,
1403
+ "mode": "affine"
1404
+ },
1405
+ "model.layers.43.mlp.switch_mlp.gate_proj": {
1406
+ "group_size": 32,
1407
+ "bits": 4,
1408
+ "mode": "mxfp4"
1409
+ },
1410
+ "model.layers.43.mlp.switch_mlp.up_proj": {
1411
+ "group_size": 32,
1412
+ "bits": 4,
1413
+ "mode": "mxfp4"
1414
+ },
1415
+ "model.layers.43.mlp.switch_mlp.down_proj": {
1416
+ "group_size": 64,
1417
+ "bits": 6,
1418
+ "mode": "affine"
1419
+ },
1420
+ "model.layers.44.linear_attn.in_proj_qkvz": {
1421
+ "group_size": 32,
1422
+ "bits": 4,
1423
+ "mode": "mxfp4"
1424
+ },
1425
+ "model.layers.44.linear_attn.in_proj_ba": {
1426
+ "group_size": 64,
1427
+ "bits": 4,
1428
+ "mode": "affine"
1429
+ },
1430
+ "model.layers.44.linear_attn.out_proj": {
1431
+ "group_size": 64,
1432
+ "bits": 6,
1433
+ "mode": "affine"
1434
+ },
1435
+ "model.layers.44.mlp.switch_mlp.gate_proj": {
1436
+ "group_size": 32,
1437
+ "bits": 4,
1438
+ "mode": "mxfp4"
1439
+ },
1440
+ "model.layers.44.mlp.switch_mlp.up_proj": {
1441
+ "group_size": 32,
1442
+ "bits": 4,
1443
+ "mode": "mxfp4"
1444
+ },
1445
+ "model.layers.44.mlp.switch_mlp.down_proj": {
1446
+ "group_size": 32,
1447
+ "bits": 4,
1448
+ "mode": "mxfp4"
1449
+ },
1450
+ "model.layers.45.linear_attn.in_proj_qkvz": {
1451
+ "group_size": 64,
1452
+ "bits": 5,
1453
+ "mode": "affine"
1454
+ },
1455
+ "model.layers.45.linear_attn.in_proj_ba": {
1456
+ "group_size": 64,
1457
+ "bits": 4,
1458
+ "mode": "affine"
1459
+ },
1460
+ "model.layers.45.linear_attn.out_proj": {
1461
+ "group_size": 64,
1462
+ "bits": 8,
1463
+ "mode": "affine"
1464
+ },
1465
+ "model.layers.45.mlp.switch_mlp.gate_proj": {
1466
+ "group_size": 32,
1467
+ "bits": 4,
1468
+ "mode": "mxfp4"
1469
+ },
1470
+ "model.layers.45.mlp.switch_mlp.up_proj": {
1471
+ "group_size": 32,
1472
+ "bits": 4,
1473
+ "mode": "mxfp4"
1474
+ },
1475
+ "model.layers.45.mlp.switch_mlp.down_proj": {
1476
+ "group_size": 64,
1477
+ "bits": 6,
1478
+ "mode": "affine"
1479
+ },
1480
+ "model.layers.46.linear_attn.in_proj_qkvz": {
1481
+ "group_size": 64,
1482
+ "bits": 5,
1483
+ "mode": "affine"
1484
+ },
1485
+ "model.layers.46.linear_attn.in_proj_ba": {
1486
+ "group_size": 64,
1487
+ "bits": 4,
1488
+ "mode": "affine"
1489
+ },
1490
+ "model.layers.46.linear_attn.out_proj": {
1491
+ "group_size": 64,
1492
+ "bits": 8,
1493
+ "mode": "affine"
1494
+ },
1495
+ "model.layers.46.mlp.switch_mlp.gate_proj": {
1496
+ "group_size": 32,
1497
+ "bits": 4,
1498
+ "mode": "mxfp4"
1499
+ },
1500
+ "model.layers.46.mlp.switch_mlp.up_proj": {
1501
+ "group_size": 32,
1502
+ "bits": 4,
1503
+ "mode": "mxfp4"
1504
+ },
1505
+ "model.layers.46.mlp.switch_mlp.down_proj": {
1506
+ "group_size": 64,
1507
+ "bits": 6,
1508
+ "mode": "affine"
1509
+ },
1510
+ "model.layers.47.self_attn.q_proj": {
1511
+ "group_size": 64,
1512
+ "bits": 5,
1513
+ "mode": "affine"
1514
+ },
1515
+ "model.layers.47.self_attn.k_proj": {
1516
+ "group_size": 64,
1517
+ "bits": 5,
1518
+ "mode": "affine"
1519
+ },
1520
+ "model.layers.47.self_attn.v_proj": {
1521
+ "group_size": 64,
1522
+ "bits": 5,
1523
+ "mode": "affine"
1524
+ },
1525
+ "model.layers.47.self_attn.o_proj": {
1526
+ "group_size": 64,
1527
+ "bits": 5,
1528
+ "mode": "affine"
1529
+ },
1530
+ "model.layers.47.mlp.switch_mlp.gate_proj": {
1531
+ "group_size": 32,
1532
+ "bits": 4,
1533
+ "mode": "mxfp4"
1534
+ },
1535
+ "model.layers.47.mlp.switch_mlp.up_proj": {
1536
+ "group_size": 32,
1537
+ "bits": 4,
1538
+ "mode": "mxfp4"
1539
+ },
1540
+ "model.layers.47.mlp.switch_mlp.down_proj": {
1541
+ "group_size": 64,
1542
+ "bits": 6,
1543
+ "mode": "affine"
1544
+ },
1545
+ "lm_head": {
1546
+ "group_size": 64,
1547
+ "bits": 6,
1548
+ "mode": "affine"
1549
+ }
1550
+ },
1551
+ "quantization_config": {
1552
+ "group_size": 64,
1553
+ "bits": 4,
1554
+ "mode": "affine",
1555
+ "model.embed_tokens": {
1556
+ "group_size": 64,
1557
+ "bits": 4,
1558
+ "mode": "affine"
1559
+ },
1560
+ "model.layers.0.linear_attn.in_proj_qkvz": {
1561
+ "group_size": 32,
1562
+ "bits": 4,
1563
+ "mode": "mxfp4"
1564
+ },
1565
+ "model.layers.0.linear_attn.in_proj_ba": {
1566
+ "group_size": 64,
1567
+ "bits": 4,
1568
+ "mode": "affine"
1569
+ },
1570
+ "model.layers.0.linear_attn.out_proj": {
1571
+ "group_size": 64,
1572
+ "bits": 5,
1573
+ "mode": "affine"
1574
+ },
1575
+ "model.layers.0.mlp.switch_mlp.gate_proj": {
1576
+ "group_size": 32,
1577
+ "bits": 4,
1578
+ "mode": "mxfp4"
1579
+ },
1580
+ "model.layers.0.mlp.switch_mlp.up_proj": {
1581
+ "group_size": 32,
1582
+ "bits": 4,
1583
+ "mode": "mxfp4"
1584
+ },
1585
+ "model.layers.0.mlp.switch_mlp.down_proj": {
1586
+ "group_size": 32,
1587
+ "bits": 4,
1588
+ "mode": "mxfp4"
1589
+ },
1590
+ "model.layers.1.linear_attn.in_proj_qkvz": {
1591
+ "group_size": 32,
1592
+ "bits": 4,
1593
+ "mode": "mxfp4"
1594
+ },
1595
+ "model.layers.1.linear_attn.in_proj_ba": {
1596
+ "group_size": 64,
1597
+ "bits": 4,
1598
+ "mode": "affine"
1599
+ },
1600
+ "model.layers.1.linear_attn.out_proj": {
1601
+ "group_size": 64,
1602
+ "bits": 5,
1603
+ "mode": "affine"
1604
+ },
1605
+ "model.layers.1.mlp.switch_mlp.gate_proj": {
1606
+ "group_size": 32,
1607
+ "bits": 4,
1608
+ "mode": "mxfp4"
1609
+ },
1610
+ "model.layers.1.mlp.switch_mlp.up_proj": {
1611
+ "group_size": 32,
1612
+ "bits": 4,
1613
+ "mode": "mxfp4"
1614
+ },
1615
+ "model.layers.1.mlp.switch_mlp.down_proj": {
1616
+ "group_size": 64,
1617
+ "bits": 6,
1618
+ "mode": "affine"
1619
+ },
1620
+ "model.layers.2.linear_attn.in_proj_qkvz": {
1621
+ "group_size": 32,
1622
+ "bits": 4,
1623
+ "mode": "mxfp4"
1624
+ },
1625
+ "model.layers.2.linear_attn.in_proj_ba": {
1626
+ "group_size": 64,
1627
+ "bits": 4,
1628
+ "mode": "affine"
1629
+ },
1630
+ "model.layers.2.linear_attn.out_proj": {
1631
+ "group_size": 64,
1632
+ "bits": 5,
1633
+ "mode": "affine"
1634
+ },
1635
+ "model.layers.2.mlp.switch_mlp.gate_proj": {
1636
+ "group_size": 32,
1637
+ "bits": 4,
1638
+ "mode": "mxfp4"
1639
+ },
1640
+ "model.layers.2.mlp.switch_mlp.up_proj": {
1641
+ "group_size": 32,
1642
+ "bits": 4,
1643
+ "mode": "mxfp4"
1644
+ },
1645
+ "model.layers.2.mlp.switch_mlp.down_proj": {
1646
+ "group_size": 32,
1647
+ "bits": 4,
1648
+ "mode": "mxfp4"
1649
+ },
1650
+ "model.layers.3.self_attn.q_proj": {
1651
+ "group_size": 32,
1652
+ "bits": 4,
1653
+ "mode": "mxfp4"
1654
+ },
1655
+ "model.layers.3.self_attn.k_proj": {
1656
+ "group_size": 32,
1657
+ "bits": 4,
1658
+ "mode": "mxfp4"
1659
+ },
1660
+ "model.layers.3.self_attn.v_proj": {
1661
+ "group_size": 32,
1662
+ "bits": 4,
1663
+ "mode": "mxfp4"
1664
+ },
1665
+ "model.layers.3.self_attn.o_proj": {
1666
+ "group_size": 32,
1667
+ "bits": 4,
1668
+ "mode": "mxfp4"
1669
+ },
1670
+ "model.layers.3.mlp.switch_mlp.gate_proj": {
1671
+ "group_size": 32,
1672
+ "bits": 4,
1673
+ "mode": "mxfp4"
1674
+ },
1675
+ "model.layers.3.mlp.switch_mlp.up_proj": {
1676
+ "group_size": 32,
1677
+ "bits": 4,
1678
+ "mode": "mxfp4"
1679
+ },
1680
+ "model.layers.3.mlp.switch_mlp.down_proj": {
1681
+ "group_size": 32,
1682
+ "bits": 4,
1683
+ "mode": "mxfp4"
1684
+ },
1685
+ "model.layers.4.linear_attn.in_proj_qkvz": {
1686
+ "group_size": 32,
1687
+ "bits": 4,
1688
+ "mode": "mxfp4"
1689
+ },
1690
+ "model.layers.4.linear_attn.in_proj_ba": {
1691
+ "group_size": 64,
1692
+ "bits": 4,
1693
+ "mode": "affine"
1694
+ },
1695
+ "model.layers.4.linear_attn.out_proj": {
1696
+ "group_size": 64,
1697
+ "bits": 5,
1698
+ "mode": "affine"
1699
+ },
1700
+ "model.layers.4.mlp.switch_mlp.gate_proj": {
1701
+ "group_size": 32,
1702
+ "bits": 4,
1703
+ "mode": "mxfp4"
1704
+ },
1705
+ "model.layers.4.mlp.switch_mlp.up_proj": {
1706
+ "group_size": 32,
1707
+ "bits": 4,
1708
+ "mode": "mxfp4"
1709
+ },
1710
+ "model.layers.4.mlp.switch_mlp.down_proj": {
1711
+ "group_size": 32,
1712
+ "bits": 4,
1713
+ "mode": "mxfp4"
1714
+ },
1715
+ "model.layers.5.linear_attn.in_proj_qkvz": {
1716
+ "group_size": 32,
1717
+ "bits": 4,
1718
+ "mode": "mxfp4"
1719
+ },
1720
+ "model.layers.5.linear_attn.in_proj_ba": {
1721
+ "group_size": 64,
1722
+ "bits": 4,
1723
+ "mode": "affine"
1724
+ },
1725
+ "model.layers.5.linear_attn.out_proj": {
1726
+ "group_size": 64,
1727
+ "bits": 5,
1728
+ "mode": "affine"
1729
+ },
1730
+ "model.layers.5.mlp.switch_mlp.gate_proj": {
1731
+ "group_size": 32,
1732
+ "bits": 4,
1733
+ "mode": "mxfp4"
1734
+ },
1735
+ "model.layers.5.mlp.switch_mlp.up_proj": {
1736
+ "group_size": 32,
1737
+ "bits": 4,
1738
+ "mode": "mxfp4"
1739
+ },
1740
+ "model.layers.5.mlp.switch_mlp.down_proj": {
1741
+ "group_size": 32,
1742
+ "bits": 4,
1743
+ "mode": "mxfp4"
1744
+ },
1745
+ "model.layers.6.linear_attn.in_proj_qkvz": {
1746
+ "group_size": 32,
1747
+ "bits": 4,
1748
+ "mode": "mxfp4"
1749
+ },
1750
+ "model.layers.6.linear_attn.in_proj_ba": {
1751
+ "group_size": 64,
1752
+ "bits": 4,
1753
+ "mode": "affine"
1754
+ },
1755
+ "model.layers.6.linear_attn.out_proj": {
1756
+ "group_size": 64,
1757
+ "bits": 5,
1758
+ "mode": "affine"
1759
+ },
1760
+ "model.layers.6.mlp.switch_mlp.gate_proj": {
1761
+ "group_size": 32,
1762
+ "bits": 4,
1763
+ "mode": "mxfp4"
1764
+ },
1765
+ "model.layers.6.mlp.switch_mlp.up_proj": {
1766
+ "group_size": 32,
1767
+ "bits": 4,
1768
+ "mode": "mxfp4"
1769
+ },
1770
+ "model.layers.6.mlp.switch_mlp.down_proj": {
1771
+ "group_size": 64,
1772
+ "bits": 6,
1773
+ "mode": "affine"
1774
+ },
1775
+ "model.layers.7.self_attn.q_proj": {
1776
+ "group_size": 32,
1777
+ "bits": 4,
1778
+ "mode": "mxfp4"
1779
+ },
1780
+ "model.layers.7.self_attn.k_proj": {
1781
+ "group_size": 32,
1782
+ "bits": 4,
1783
+ "mode": "mxfp4"
1784
+ },
1785
+ "model.layers.7.self_attn.v_proj": {
1786
+ "group_size": 32,
1787
+ "bits": 4,
1788
+ "mode": "mxfp4"
1789
+ },
1790
+ "model.layers.7.self_attn.o_proj": {
1791
+ "group_size": 32,
1792
+ "bits": 4,
1793
+ "mode": "mxfp4"
1794
+ },
1795
+ "model.layers.7.mlp.switch_mlp.gate_proj": {
1796
+ "group_size": 32,
1797
+ "bits": 4,
1798
+ "mode": "mxfp4"
1799
+ },
1800
+ "model.layers.7.mlp.switch_mlp.up_proj": {
1801
+ "group_size": 32,
1802
+ "bits": 4,
1803
+ "mode": "mxfp4"
1804
+ },
1805
+ "model.layers.7.mlp.switch_mlp.down_proj": {
1806
+ "group_size": 32,
1807
+ "bits": 4,
1808
+ "mode": "mxfp4"
1809
+ },
1810
+ "model.layers.8.linear_attn.in_proj_qkvz": {
1811
+ "group_size": 32,
1812
+ "bits": 4,
1813
+ "mode": "mxfp4"
1814
+ },
1815
+ "model.layers.8.linear_attn.in_proj_ba": {
1816
+ "group_size": 64,
1817
+ "bits": 4,
1818
+ "mode": "affine"
1819
+ },
1820
+ "model.layers.8.linear_attn.out_proj": {
1821
+ "group_size": 64,
1822
+ "bits": 5,
1823
+ "mode": "affine"
1824
+ },
1825
+ "model.layers.8.mlp.switch_mlp.gate_proj": {
1826
+ "group_size": 32,
1827
+ "bits": 4,
1828
+ "mode": "mxfp4"
1829
+ },
1830
+ "model.layers.8.mlp.switch_mlp.up_proj": {
1831
+ "group_size": 32,
1832
+ "bits": 4,
1833
+ "mode": "mxfp4"
1834
+ },
1835
+ "model.layers.8.mlp.switch_mlp.down_proj": {
1836
+ "group_size": 32,
1837
+ "bits": 4,
1838
+ "mode": "mxfp4"
1839
+ },
1840
+ "model.layers.9.linear_attn.in_proj_qkvz": {
1841
+ "group_size": 32,
1842
+ "bits": 4,
1843
+ "mode": "mxfp4"
1844
+ },
1845
+ "model.layers.9.linear_attn.in_proj_ba": {
1846
+ "group_size": 64,
1847
+ "bits": 4,
1848
+ "mode": "affine"
1849
+ },
1850
+ "model.layers.9.linear_attn.out_proj": {
1851
+ "group_size": 64,
1852
+ "bits": 5,
1853
+ "mode": "affine"
1854
+ },
1855
+ "model.layers.9.mlp.switch_mlp.gate_proj": {
1856
+ "group_size": 32,
1857
+ "bits": 4,
1858
+ "mode": "mxfp4"
1859
+ },
1860
+ "model.layers.9.mlp.switch_mlp.up_proj": {
1861
+ "group_size": 32,
1862
+ "bits": 4,
1863
+ "mode": "mxfp4"
1864
+ },
1865
+ "model.layers.9.mlp.switch_mlp.down_proj": {
1866
+ "group_size": 32,
1867
+ "bits": 4,
1868
+ "mode": "mxfp4"
1869
+ },
1870
+ "model.layers.10.linear_attn.in_proj_qkvz": {
1871
+ "group_size": 32,
1872
+ "bits": 4,
1873
+ "mode": "mxfp4"
1874
+ },
1875
+ "model.layers.10.linear_attn.in_proj_ba": {
1876
+ "group_size": 64,
1877
+ "bits": 4,
1878
+ "mode": "affine"
1879
+ },
1880
+ "model.layers.10.linear_attn.out_proj": {
1881
+ "group_size": 64,
1882
+ "bits": 5,
1883
+ "mode": "affine"
1884
+ },
1885
+ "model.layers.10.mlp.switch_mlp.gate_proj": {
1886
+ "group_size": 32,
1887
+ "bits": 4,
1888
+ "mode": "mxfp4"
1889
+ },
1890
+ "model.layers.10.mlp.switch_mlp.up_proj": {
1891
+ "group_size": 32,
1892
+ "bits": 4,
1893
+ "mode": "mxfp4"
1894
+ },
1895
+ "model.layers.10.mlp.switch_mlp.down_proj": {
1896
+ "group_size": 64,
1897
+ "bits": 6,
1898
+ "mode": "affine"
1899
+ },
1900
+ "model.layers.11.self_attn.q_proj": {
1901
+ "group_size": 32,
1902
+ "bits": 4,
1903
+ "mode": "mxfp4"
1904
+ },
1905
+ "model.layers.11.self_attn.k_proj": {
1906
+ "group_size": 32,
1907
+ "bits": 4,
1908
+ "mode": "mxfp4"
1909
+ },
1910
+ "model.layers.11.self_attn.v_proj": {
1911
+ "group_size": 32,
1912
+ "bits": 4,
1913
+ "mode": "mxfp4"
1914
+ },
1915
+ "model.layers.11.self_attn.o_proj": {
1916
+ "group_size": 32,
1917
+ "bits": 4,
1918
+ "mode": "mxfp4"
1919
+ },
1920
+ "model.layers.11.mlp.switch_mlp.gate_proj": {
1921
+ "group_size": 32,
1922
+ "bits": 4,
1923
+ "mode": "mxfp4"
1924
+ },
1925
+ "model.layers.11.mlp.switch_mlp.up_proj": {
1926
+ "group_size": 32,
1927
+ "bits": 4,
1928
+ "mode": "mxfp4"
1929
+ },
1930
+ "model.layers.11.mlp.switch_mlp.down_proj": {
1931
+ "group_size": 32,
1932
+ "bits": 4,
1933
+ "mode": "mxfp4"
1934
+ },
1935
+ "model.layers.12.linear_attn.in_proj_qkvz": {
1936
+ "group_size": 32,
1937
+ "bits": 4,
1938
+ "mode": "mxfp4"
1939
+ },
1940
+ "model.layers.12.linear_attn.in_proj_ba": {
1941
+ "group_size": 64,
1942
+ "bits": 4,
1943
+ "mode": "affine"
1944
+ },
1945
+ "model.layers.12.linear_attn.out_proj": {
1946
+ "group_size": 64,
1947
+ "bits": 6,
1948
+ "mode": "affine"
1949
+ },
1950
+ "model.layers.12.mlp.switch_mlp.gate_proj": {
1951
+ "group_size": 32,
1952
+ "bits": 4,
1953
+ "mode": "mxfp4"
1954
+ },
1955
+ "model.layers.12.mlp.switch_mlp.up_proj": {
1956
+ "group_size": 32,
1957
+ "bits": 4,
1958
+ "mode": "mxfp4"
1959
+ },
1960
+ "model.layers.12.mlp.switch_mlp.down_proj": {
1961
+ "group_size": 32,
1962
+ "bits": 4,
1963
+ "mode": "mxfp4"
1964
+ },
1965
+ "model.layers.13.linear_attn.in_proj_qkvz": {
1966
+ "group_size": 32,
1967
+ "bits": 4,
1968
+ "mode": "mxfp4"
1969
+ },
1970
+ "model.layers.13.linear_attn.in_proj_ba": {
1971
+ "group_size": 64,
1972
+ "bits": 4,
1973
+ "mode": "affine"
1974
+ },
1975
+ "model.layers.13.linear_attn.out_proj": {
1976
+ "group_size": 64,
1977
+ "bits": 5,
1978
+ "mode": "affine"
1979
+ },
1980
+ "model.layers.13.mlp.switch_mlp.gate_proj": {
1981
+ "group_size": 32,
1982
+ "bits": 4,
1983
+ "mode": "mxfp4"
1984
+ },
1985
+ "model.layers.13.mlp.switch_mlp.up_proj": {
1986
+ "group_size": 32,
1987
+ "bits": 4,
1988
+ "mode": "mxfp4"
1989
+ },
1990
+ "model.layers.13.mlp.switch_mlp.down_proj": {
1991
+ "group_size": 32,
1992
+ "bits": 4,
1993
+ "mode": "mxfp4"
1994
+ },
1995
+ "model.layers.14.linear_attn.in_proj_qkvz": {
1996
+ "group_size": 32,
1997
+ "bits": 4,
1998
+ "mode": "mxfp4"
1999
+ },
2000
+ "model.layers.14.linear_attn.in_proj_ba": {
2001
+ "group_size": 64,
2002
+ "bits": 4,
2003
+ "mode": "affine"
2004
+ },
2005
+ "model.layers.14.linear_attn.out_proj": {
2006
+ "group_size": 64,
2007
+ "bits": 5,
2008
+ "mode": "affine"
2009
+ },
2010
+ "model.layers.14.mlp.switch_mlp.gate_proj": {
2011
+ "group_size": 32,
2012
+ "bits": 4,
2013
+ "mode": "mxfp4"
2014
+ },
2015
+ "model.layers.14.mlp.switch_mlp.up_proj": {
2016
+ "group_size": 32,
2017
+ "bits": 4,
2018
+ "mode": "mxfp4"
2019
+ },
2020
+ "model.layers.14.mlp.switch_mlp.down_proj": {
2021
+ "group_size": 64,
2022
+ "bits": 6,
2023
+ "mode": "affine"
2024
+ },
2025
+ "model.layers.15.self_attn.q_proj": {
2026
+ "group_size": 32,
2027
+ "bits": 4,
2028
+ "mode": "mxfp4"
2029
+ },
2030
+ "model.layers.15.self_attn.k_proj": {
2031
+ "group_size": 32,
2032
+ "bits": 4,
2033
+ "mode": "mxfp4"
2034
+ },
2035
+ "model.layers.15.self_attn.v_proj": {
2036
+ "group_size": 32,
2037
+ "bits": 4,
2038
+ "mode": "mxfp4"
2039
+ },
2040
+ "model.layers.15.self_attn.o_proj": {
2041
+ "group_size": 32,
2042
+ "bits": 4,
2043
+ "mode": "mxfp4"
2044
+ },
2045
+ "model.layers.15.mlp.switch_mlp.gate_proj": {
2046
+ "group_size": 32,
2047
+ "bits": 4,
2048
+ "mode": "mxfp4"
2049
+ },
2050
+ "model.layers.15.mlp.switch_mlp.up_proj": {
2051
+ "group_size": 32,
2052
+ "bits": 4,
2053
+ "mode": "mxfp4"
2054
+ },
2055
+ "model.layers.15.mlp.switch_mlp.down_proj": {
2056
+ "group_size": 32,
2057
+ "bits": 4,
2058
+ "mode": "mxfp4"
2059
+ },
2060
+ "model.layers.16.linear_attn.in_proj_qkvz": {
2061
+ "group_size": 32,
2062
+ "bits": 4,
2063
+ "mode": "mxfp4"
2064
+ },
2065
+ "model.layers.16.linear_attn.in_proj_ba": {
2066
+ "group_size": 64,
2067
+ "bits": 4,
2068
+ "mode": "affine"
2069
+ },
2070
+ "model.layers.16.linear_attn.out_proj": {
2071
+ "group_size": 64,
2072
+ "bits": 6,
2073
+ "mode": "affine"
2074
+ },
2075
+ "model.layers.16.mlp.switch_mlp.gate_proj": {
2076
+ "group_size": 32,
2077
+ "bits": 4,
2078
+ "mode": "mxfp4"
2079
+ },
2080
+ "model.layers.16.mlp.switch_mlp.up_proj": {
2081
+ "group_size": 32,
2082
+ "bits": 4,
2083
+ "mode": "mxfp4"
2084
+ },
2085
+ "model.layers.16.mlp.switch_mlp.down_proj": {
2086
+ "group_size": 32,
2087
+ "bits": 4,
2088
+ "mode": "mxfp4"
2089
+ },
2090
+ "model.layers.17.linear_attn.in_proj_qkvz": {
2091
+ "group_size": 32,
2092
+ "bits": 4,
2093
+ "mode": "mxfp4"
2094
+ },
2095
+ "model.layers.17.linear_attn.in_proj_ba": {
2096
+ "group_size": 64,
2097
+ "bits": 4,
2098
+ "mode": "affine"
2099
+ },
2100
+ "model.layers.17.linear_attn.out_proj": {
2101
+ "group_size": 64,
2102
+ "bits": 6,
2103
+ "mode": "affine"
2104
+ },
2105
+ "model.layers.17.mlp.switch_mlp.gate_proj": {
2106
+ "group_size": 32,
2107
+ "bits": 4,
2108
+ "mode": "mxfp4"
2109
+ },
2110
+ "model.layers.17.mlp.switch_mlp.up_proj": {
2111
+ "group_size": 32,
2112
+ "bits": 4,
2113
+ "mode": "mxfp4"
2114
+ },
2115
+ "model.layers.17.mlp.switch_mlp.down_proj": {
2116
+ "group_size": 32,
2117
+ "bits": 4,
2118
+ "mode": "mxfp4"
2119
+ },
2120
+ "model.layers.18.linear_attn.in_proj_qkvz": {
2121
+ "group_size": 32,
2122
+ "bits": 4,
2123
+ "mode": "mxfp4"
2124
+ },
2125
+ "model.layers.18.linear_attn.in_proj_ba": {
2126
+ "group_size": 64,
2127
+ "bits": 4,
2128
+ "mode": "affine"
2129
+ },
2130
+ "model.layers.18.linear_attn.out_proj": {
2131
+ "group_size": 64,
2132
+ "bits": 5,
2133
+ "mode": "affine"
2134
+ },
2135
+ "model.layers.18.mlp.switch_mlp.gate_proj": {
2136
+ "group_size": 32,
2137
+ "bits": 4,
2138
+ "mode": "mxfp4"
2139
+ },
2140
+ "model.layers.18.mlp.switch_mlp.up_proj": {
2141
+ "group_size": 32,
2142
+ "bits": 4,
2143
+ "mode": "mxfp4"
2144
+ },
2145
+ "model.layers.18.mlp.switch_mlp.down_proj": {
2146
+ "group_size": 32,
2147
+ "bits": 4,
2148
+ "mode": "mxfp4"
2149
+ },
2150
+ "model.layers.19.self_attn.q_proj": {
2151
+ "group_size": 32,
2152
+ "bits": 4,
2153
+ "mode": "mxfp4"
2154
+ },
2155
+ "model.layers.19.self_attn.k_proj": {
2156
+ "group_size": 32,
2157
+ "bits": 4,
2158
+ "mode": "mxfp4"
2159
+ },
2160
+ "model.layers.19.self_attn.v_proj": {
2161
+ "group_size": 32,
2162
+ "bits": 4,
2163
+ "mode": "mxfp4"
2164
+ },
2165
+ "model.layers.19.self_attn.o_proj": {
2166
+ "group_size": 32,
2167
+ "bits": 4,
2168
+ "mode": "mxfp4"
2169
+ },
2170
+ "model.layers.19.mlp.switch_mlp.gate_proj": {
2171
+ "group_size": 32,
2172
+ "bits": 4,
2173
+ "mode": "mxfp4"
2174
+ },
2175
+ "model.layers.19.mlp.switch_mlp.up_proj": {
2176
+ "group_size": 32,
2177
+ "bits": 4,
2178
+ "mode": "mxfp4"
2179
+ },
2180
+ "model.layers.19.mlp.switch_mlp.down_proj": {
2181
+ "group_size": 32,
2182
+ "bits": 4,
2183
+ "mode": "mxfp4"
2184
+ },
2185
+ "model.layers.20.linear_attn.in_proj_qkvz": {
2186
+ "group_size": 32,
2187
+ "bits": 4,
2188
+ "mode": "mxfp4"
2189
+ },
2190
+ "model.layers.20.linear_attn.in_proj_ba": {
2191
+ "group_size": 64,
2192
+ "bits": 4,
2193
+ "mode": "affine"
2194
+ },
2195
+ "model.layers.20.linear_attn.out_proj": {
2196
+ "group_size": 64,
2197
+ "bits": 5,
2198
+ "mode": "affine"
2199
+ },
2200
+ "model.layers.20.mlp.switch_mlp.gate_proj": {
2201
+ "group_size": 32,
2202
+ "bits": 4,
2203
+ "mode": "mxfp4"
2204
+ },
2205
+ "model.layers.20.mlp.switch_mlp.up_proj": {
2206
+ "group_size": 32,
2207
+ "bits": 4,
2208
+ "mode": "mxfp4"
2209
+ },
2210
+ "model.layers.20.mlp.switch_mlp.down_proj": {
2211
+ "group_size": 32,
2212
+ "bits": 4,
2213
+ "mode": "mxfp4"
2214
+ },
2215
+ "model.layers.21.linear_attn.in_proj_qkvz": {
2216
+ "group_size": 32,
2217
+ "bits": 4,
2218
+ "mode": "mxfp4"
2219
+ },
2220
+ "model.layers.21.linear_attn.in_proj_ba": {
2221
+ "group_size": 64,
2222
+ "bits": 4,
2223
+ "mode": "affine"
2224
+ },
2225
+ "model.layers.21.linear_attn.out_proj": {
2226
+ "group_size": 64,
2227
+ "bits": 5,
2228
+ "mode": "affine"
2229
+ },
2230
+ "model.layers.21.mlp.switch_mlp.gate_proj": {
2231
+ "group_size": 32,
2232
+ "bits": 4,
2233
+ "mode": "mxfp4"
2234
+ },
2235
+ "model.layers.21.mlp.switch_mlp.up_proj": {
2236
+ "group_size": 32,
2237
+ "bits": 4,
2238
+ "mode": "mxfp4"
2239
+ },
2240
+ "model.layers.21.mlp.switch_mlp.down_proj": {
2241
+ "group_size": 32,
2242
+ "bits": 4,
2243
+ "mode": "mxfp4"
2244
+ },
2245
+ "model.layers.22.linear_attn.in_proj_qkvz": {
2246
+ "group_size": 32,
2247
+ "bits": 4,
2248
+ "mode": "mxfp4"
2249
+ },
2250
+ "model.layers.22.linear_attn.in_proj_ba": {
2251
+ "group_size": 64,
2252
+ "bits": 4,
2253
+ "mode": "affine"
2254
+ },
2255
+ "model.layers.22.linear_attn.out_proj": {
2256
+ "group_size": 64,
2257
+ "bits": 5,
2258
+ "mode": "affine"
2259
+ },
2260
+ "model.layers.22.mlp.switch_mlp.gate_proj": {
2261
+ "group_size": 32,
2262
+ "bits": 4,
2263
+ "mode": "mxfp4"
2264
+ },
2265
+ "model.layers.22.mlp.switch_mlp.up_proj": {
2266
+ "group_size": 32,
2267
+ "bits": 4,
2268
+ "mode": "mxfp4"
2269
+ },
2270
+ "model.layers.22.mlp.switch_mlp.down_proj": {
2271
+ "group_size": 64,
2272
+ "bits": 6,
2273
+ "mode": "affine"
2274
+ },
2275
+ "model.layers.23.self_attn.q_proj": {
2276
+ "group_size": 32,
2277
+ "bits": 4,
2278
+ "mode": "mxfp4"
2279
+ },
2280
+ "model.layers.23.self_attn.k_proj": {
2281
+ "group_size": 32,
2282
+ "bits": 4,
2283
+ "mode": "mxfp4"
2284
+ },
2285
+ "model.layers.23.self_attn.v_proj": {
2286
+ "group_size": 32,
2287
+ "bits": 4,
2288
+ "mode": "mxfp4"
2289
+ },
2290
+ "model.layers.23.self_attn.o_proj": {
2291
+ "group_size": 32,
2292
+ "bits": 4,
2293
+ "mode": "mxfp4"
2294
+ },
2295
+ "model.layers.23.mlp.switch_mlp.gate_proj": {
2296
+ "group_size": 32,
2297
+ "bits": 4,
2298
+ "mode": "mxfp4"
2299
+ },
2300
+ "model.layers.23.mlp.switch_mlp.up_proj": {
2301
+ "group_size": 32,
2302
+ "bits": 4,
2303
+ "mode": "mxfp4"
2304
+ },
2305
+ "model.layers.23.mlp.switch_mlp.down_proj": {
2306
+ "group_size": 32,
2307
+ "bits": 4,
2308
+ "mode": "mxfp4"
2309
+ },
2310
+ "model.layers.24.linear_attn.in_proj_qkvz": {
2311
+ "group_size": 32,
2312
+ "bits": 4,
2313
+ "mode": "mxfp4"
2314
+ },
2315
+ "model.layers.24.linear_attn.in_proj_ba": {
2316
+ "group_size": 64,
2317
+ "bits": 4,
2318
+ "mode": "affine"
2319
+ },
2320
+ "model.layers.24.linear_attn.out_proj": {
2321
+ "group_size": 64,
2322
+ "bits": 5,
2323
+ "mode": "affine"
2324
+ },
2325
+ "model.layers.24.mlp.switch_mlp.gate_proj": {
2326
+ "group_size": 32,
2327
+ "bits": 4,
2328
+ "mode": "mxfp4"
2329
+ },
2330
+ "model.layers.24.mlp.switch_mlp.up_proj": {
2331
+ "group_size": 32,
2332
+ "bits": 4,
2333
+ "mode": "mxfp4"
2334
+ },
2335
+ "model.layers.24.mlp.switch_mlp.down_proj": {
2336
+ "group_size": 32,
2337
+ "bits": 4,
2338
+ "mode": "mxfp4"
2339
+ },
2340
+ "model.layers.25.linear_attn.in_proj_qkvz": {
2341
+ "group_size": 32,
2342
+ "bits": 4,
2343
+ "mode": "mxfp4"
2344
+ },
2345
+ "model.layers.25.linear_attn.in_proj_ba": {
2346
+ "group_size": 64,
2347
+ "bits": 4,
2348
+ "mode": "affine"
2349
+ },
2350
+ "model.layers.25.linear_attn.out_proj": {
2351
+ "group_size": 64,
2352
+ "bits": 5,
2353
+ "mode": "affine"
2354
+ },
2355
+ "model.layers.25.mlp.switch_mlp.gate_proj": {
2356
+ "group_size": 32,
2357
+ "bits": 4,
2358
+ "mode": "mxfp4"
2359
+ },
2360
+ "model.layers.25.mlp.switch_mlp.up_proj": {
2361
+ "group_size": 32,
2362
+ "bits": 4,
2363
+ "mode": "mxfp4"
2364
+ },
2365
+ "model.layers.25.mlp.switch_mlp.down_proj": {
2366
+ "group_size": 32,
2367
+ "bits": 4,
2368
+ "mode": "mxfp4"
2369
+ },
2370
+ "model.layers.26.linear_attn.in_proj_qkvz": {
2371
+ "group_size": 32,
2372
+ "bits": 4,
2373
+ "mode": "mxfp4"
2374
+ },
2375
+ "model.layers.26.linear_attn.in_proj_ba": {
2376
+ "group_size": 64,
2377
+ "bits": 4,
2378
+ "mode": "affine"
2379
+ },
2380
+ "model.layers.26.linear_attn.out_proj": {
2381
+ "group_size": 64,
2382
+ "bits": 5,
2383
+ "mode": "affine"
2384
+ },
2385
+ "model.layers.26.mlp.switch_mlp.gate_proj": {
2386
+ "group_size": 32,
2387
+ "bits": 4,
2388
+ "mode": "mxfp4"
2389
+ },
2390
+ "model.layers.26.mlp.switch_mlp.up_proj": {
2391
+ "group_size": 32,
2392
+ "bits": 4,
2393
+ "mode": "mxfp4"
2394
+ },
2395
+ "model.layers.26.mlp.switch_mlp.down_proj": {
2396
+ "group_size": 32,
2397
+ "bits": 4,
2398
+ "mode": "mxfp4"
2399
+ },
2400
+ "model.layers.27.self_attn.q_proj": {
2401
+ "group_size": 32,
2402
+ "bits": 4,
2403
+ "mode": "mxfp4"
2404
+ },
2405
+ "model.layers.27.self_attn.k_proj": {
2406
+ "group_size": 32,
2407
+ "bits": 4,
2408
+ "mode": "mxfp4"
2409
+ },
2410
+ "model.layers.27.self_attn.v_proj": {
2411
+ "group_size": 32,
2412
+ "bits": 4,
2413
+ "mode": "mxfp4"
2414
+ },
2415
+ "model.layers.27.self_attn.o_proj": {
2416
+ "group_size": 32,
2417
+ "bits": 4,
2418
+ "mode": "mxfp4"
2419
+ },
2420
+ "model.layers.27.mlp.switch_mlp.gate_proj": {
2421
+ "group_size": 32,
2422
+ "bits": 4,
2423
+ "mode": "mxfp4"
2424
+ },
2425
+ "model.layers.27.mlp.switch_mlp.up_proj": {
2426
+ "group_size": 32,
2427
+ "bits": 4,
2428
+ "mode": "mxfp4"
2429
+ },
2430
+ "model.layers.27.mlp.switch_mlp.down_proj": {
2431
+ "group_size": 32,
2432
+ "bits": 4,
2433
+ "mode": "mxfp4"
2434
+ },
2435
+ "model.layers.28.linear_attn.in_proj_qkvz": {
2436
+ "group_size": 32,
2437
+ "bits": 4,
2438
+ "mode": "mxfp4"
2439
+ },
2440
+ "model.layers.28.linear_attn.in_proj_ba": {
2441
+ "group_size": 64,
2442
+ "bits": 4,
2443
+ "mode": "affine"
2444
+ },
2445
+ "model.layers.28.linear_attn.out_proj": {
2446
+ "group_size": 64,
2447
+ "bits": 5,
2448
+ "mode": "affine"
2449
+ },
2450
+ "model.layers.28.mlp.switch_mlp.gate_proj": {
2451
+ "group_size": 32,
2452
+ "bits": 4,
2453
+ "mode": "mxfp4"
2454
+ },
2455
+ "model.layers.28.mlp.switch_mlp.up_proj": {
2456
+ "group_size": 32,
2457
+ "bits": 4,
2458
+ "mode": "mxfp4"
2459
+ },
2460
+ "model.layers.28.mlp.switch_mlp.down_proj": {
2461
+ "group_size": 32,
2462
+ "bits": 4,
2463
+ "mode": "mxfp4"
2464
+ },
2465
+ "model.layers.29.linear_attn.in_proj_qkvz": {
2466
+ "group_size": 32,
2467
+ "bits": 4,
2468
+ "mode": "mxfp4"
2469
+ },
2470
+ "model.layers.29.linear_attn.in_proj_ba": {
2471
+ "group_size": 64,
2472
+ "bits": 4,
2473
+ "mode": "affine"
2474
+ },
2475
+ "model.layers.29.linear_attn.out_proj": {
2476
+ "group_size": 64,
2477
+ "bits": 5,
2478
+ "mode": "affine"
2479
+ },
2480
+ "model.layers.29.mlp.switch_mlp.gate_proj": {
2481
+ "group_size": 32,
2482
+ "bits": 4,
2483
+ "mode": "mxfp4"
2484
+ },
2485
+ "model.layers.29.mlp.switch_mlp.up_proj": {
2486
+ "group_size": 32,
2487
+ "bits": 4,
2488
+ "mode": "mxfp4"
2489
+ },
2490
+ "model.layers.29.mlp.switch_mlp.down_proj": {
2491
+ "group_size": 32,
2492
+ "bits": 4,
2493
+ "mode": "mxfp4"
2494
+ },
2495
+ "model.layers.30.linear_attn.in_proj_qkvz": {
2496
+ "group_size": 32,
2497
+ "bits": 4,
2498
+ "mode": "mxfp4"
2499
+ },
2500
+ "model.layers.30.linear_attn.in_proj_ba": {
2501
+ "group_size": 64,
2502
+ "bits": 4,
2503
+ "mode": "affine"
2504
+ },
2505
+ "model.layers.30.linear_attn.out_proj": {
2506
+ "group_size": 64,
2507
+ "bits": 6,
2508
+ "mode": "affine"
2509
+ },
2510
+ "model.layers.30.mlp.switch_mlp.gate_proj": {
2511
+ "group_size": 32,
2512
+ "bits": 4,
2513
+ "mode": "mxfp4"
2514
+ },
2515
+ "model.layers.30.mlp.switch_mlp.up_proj": {
2516
+ "group_size": 32,
2517
+ "bits": 4,
2518
+ "mode": "mxfp4"
2519
+ },
2520
+ "model.layers.30.mlp.switch_mlp.down_proj": {
2521
+ "group_size": 32,
2522
+ "bits": 4,
2523
+ "mode": "mxfp4"
2524
+ },
2525
+ "model.layers.31.self_attn.q_proj": {
2526
+ "group_size": 32,
2527
+ "bits": 4,
2528
+ "mode": "mxfp4"
2529
+ },
2530
+ "model.layers.31.self_attn.k_proj": {
2531
+ "group_size": 32,
2532
+ "bits": 4,
2533
+ "mode": "mxfp4"
2534
+ },
2535
+ "model.layers.31.self_attn.v_proj": {
2536
+ "group_size": 32,
2537
+ "bits": 4,
2538
+ "mode": "mxfp4"
2539
+ },
2540
+ "model.layers.31.self_attn.o_proj": {
2541
+ "group_size": 32,
2542
+ "bits": 4,
2543
+ "mode": "mxfp4"
2544
+ },
2545
+ "model.layers.31.mlp.switch_mlp.gate_proj": {
2546
+ "group_size": 32,
2547
+ "bits": 4,
2548
+ "mode": "mxfp4"
2549
+ },
2550
+ "model.layers.31.mlp.switch_mlp.up_proj": {
2551
+ "group_size": 32,
2552
+ "bits": 4,
2553
+ "mode": "mxfp4"
2554
+ },
2555
+ "model.layers.31.mlp.switch_mlp.down_proj": {
2556
+ "group_size": 64,
2557
+ "bits": 6,
2558
+ "mode": "affine"
2559
+ },
2560
+ "model.layers.32.linear_attn.in_proj_qkvz": {
2561
+ "group_size": 32,
2562
+ "bits": 4,
2563
+ "mode": "mxfp4"
2564
+ },
2565
+ "model.layers.32.linear_attn.in_proj_ba": {
2566
+ "group_size": 64,
2567
+ "bits": 4,
2568
+ "mode": "affine"
2569
+ },
2570
+ "model.layers.32.linear_attn.out_proj": {
2571
+ "group_size": 64,
2572
+ "bits": 6,
2573
+ "mode": "affine"
2574
+ },
2575
+ "model.layers.32.mlp.switch_mlp.gate_proj": {
2576
+ "group_size": 32,
2577
+ "bits": 4,
2578
+ "mode": "mxfp4"
2579
+ },
2580
+ "model.layers.32.mlp.switch_mlp.up_proj": {
2581
+ "group_size": 32,
2582
+ "bits": 4,
2583
+ "mode": "mxfp4"
2584
+ },
2585
+ "model.layers.32.mlp.switch_mlp.down_proj": {
2586
+ "group_size": 32,
2587
+ "bits": 4,
2588
+ "mode": "mxfp4"
2589
+ },
2590
+ "model.layers.33.linear_attn.in_proj_qkvz": {
2591
+ "group_size": 32,
2592
+ "bits": 4,
2593
+ "mode": "mxfp4"
2594
+ },
2595
+ "model.layers.33.linear_attn.in_proj_ba": {
2596
+ "group_size": 64,
2597
+ "bits": 4,
2598
+ "mode": "affine"
2599
+ },
2600
+ "model.layers.33.linear_attn.out_proj": {
2601
+ "group_size": 64,
2602
+ "bits": 5,
2603
+ "mode": "affine"
2604
+ },
2605
+ "model.layers.33.mlp.switch_mlp.gate_proj": {
2606
+ "group_size": 32,
2607
+ "bits": 4,
2608
+ "mode": "mxfp4"
2609
+ },
2610
+ "model.layers.33.mlp.switch_mlp.up_proj": {
2611
+ "group_size": 32,
2612
+ "bits": 4,
2613
+ "mode": "mxfp4"
2614
+ },
2615
+ "model.layers.33.mlp.switch_mlp.down_proj": {
2616
+ "group_size": 32,
2617
+ "bits": 4,
2618
+ "mode": "mxfp4"
2619
+ },
2620
+ "model.layers.34.linear_attn.in_proj_qkvz": {
2621
+ "group_size": 32,
2622
+ "bits": 4,
2623
+ "mode": "mxfp4"
2624
+ },
2625
+ "model.layers.34.linear_attn.in_proj_ba": {
2626
+ "group_size": 64,
2627
+ "bits": 4,
2628
+ "mode": "affine"
2629
+ },
2630
+ "model.layers.34.linear_attn.out_proj": {
2631
+ "group_size": 64,
2632
+ "bits": 6,
2633
+ "mode": "affine"
2634
+ },
2635
+ "model.layers.34.mlp.switch_mlp.gate_proj": {
2636
+ "group_size": 32,
2637
+ "bits": 4,
2638
+ "mode": "mxfp4"
2639
+ },
2640
+ "model.layers.34.mlp.switch_mlp.up_proj": {
2641
+ "group_size": 32,
2642
+ "bits": 4,
2643
+ "mode": "mxfp4"
2644
+ },
2645
+ "model.layers.34.mlp.switch_mlp.down_proj": {
2646
+ "group_size": 32,
2647
+ "bits": 4,
2648
+ "mode": "mxfp4"
2649
+ },
2650
+ "model.layers.35.self_attn.q_proj": {
2651
+ "group_size": 32,
2652
+ "bits": 4,
2653
+ "mode": "mxfp4"
2654
+ },
2655
+ "model.layers.35.self_attn.k_proj": {
2656
+ "group_size": 32,
2657
+ "bits": 4,
2658
+ "mode": "mxfp4"
2659
+ },
2660
+ "model.layers.35.self_attn.v_proj": {
2661
+ "group_size": 32,
2662
+ "bits": 4,
2663
+ "mode": "mxfp4"
2664
+ },
2665
+ "model.layers.35.self_attn.o_proj": {
2666
+ "group_size": 32,
2667
+ "bits": 4,
2668
+ "mode": "mxfp4"
2669
+ },
2670
+ "model.layers.35.mlp.switch_mlp.gate_proj": {
2671
+ "group_size": 32,
2672
+ "bits": 4,
2673
+ "mode": "mxfp4"
2674
+ },
2675
+ "model.layers.35.mlp.switch_mlp.up_proj": {
2676
+ "group_size": 32,
2677
+ "bits": 4,
2678
+ "mode": "mxfp4"
2679
+ },
2680
+ "model.layers.35.mlp.switch_mlp.down_proj": {
2681
+ "group_size": 32,
2682
+ "bits": 4,
2683
+ "mode": "mxfp4"
2684
+ },
2685
+ "model.layers.36.linear_attn.in_proj_qkvz": {
2686
+ "group_size": 32,
2687
+ "bits": 4,
2688
+ "mode": "mxfp4"
2689
+ },
2690
+ "model.layers.36.linear_attn.in_proj_ba": {
2691
+ "group_size": 64,
2692
+ "bits": 4,
2693
+ "mode": "affine"
2694
+ },
2695
+ "model.layers.36.linear_attn.out_proj": {
2696
+ "group_size": 64,
2697
+ "bits": 5,
2698
+ "mode": "affine"
2699
+ },
2700
+ "model.layers.36.mlp.switch_mlp.gate_proj": {
2701
+ "group_size": 32,
2702
+ "bits": 4,
2703
+ "mode": "mxfp4"
2704
+ },
2705
+ "model.layers.36.mlp.switch_mlp.up_proj": {
2706
+ "group_size": 32,
2707
+ "bits": 4,
2708
+ "mode": "mxfp4"
2709
+ },
2710
+ "model.layers.36.mlp.switch_mlp.down_proj": {
2711
+ "group_size": 32,
2712
+ "bits": 4,
2713
+ "mode": "mxfp4"
2714
+ },
2715
+ "model.layers.37.linear_attn.in_proj_qkvz": {
2716
+ "group_size": 32,
2717
+ "bits": 4,
2718
+ "mode": "mxfp4"
2719
+ },
2720
+ "model.layers.37.linear_attn.in_proj_ba": {
2721
+ "group_size": 64,
2722
+ "bits": 4,
2723
+ "mode": "affine"
2724
+ },
2725
+ "model.layers.37.linear_attn.out_proj": {
2726
+ "group_size": 64,
2727
+ "bits": 5,
2728
+ "mode": "affine"
2729
+ },
2730
+ "model.layers.37.mlp.switch_mlp.gate_proj": {
2731
+ "group_size": 32,
2732
+ "bits": 4,
2733
+ "mode": "mxfp4"
2734
+ },
2735
+ "model.layers.37.mlp.switch_mlp.up_proj": {
2736
+ "group_size": 32,
2737
+ "bits": 4,
2738
+ "mode": "mxfp4"
2739
+ },
2740
+ "model.layers.37.mlp.switch_mlp.down_proj": {
2741
+ "group_size": 32,
2742
+ "bits": 4,
2743
+ "mode": "mxfp4"
2744
+ },
2745
+ "model.layers.38.linear_attn.in_proj_qkvz": {
2746
+ "group_size": 32,
2747
+ "bits": 4,
2748
+ "mode": "mxfp4"
2749
+ },
2750
+ "model.layers.38.linear_attn.in_proj_ba": {
2751
+ "group_size": 64,
2752
+ "bits": 4,
2753
+ "mode": "affine"
2754
+ },
2755
+ "model.layers.38.linear_attn.out_proj": {
2756
+ "group_size": 64,
2757
+ "bits": 5,
2758
+ "mode": "affine"
2759
+ },
2760
+ "model.layers.38.mlp.switch_mlp.gate_proj": {
2761
+ "group_size": 32,
2762
+ "bits": 4,
2763
+ "mode": "mxfp4"
2764
+ },
2765
+ "model.layers.38.mlp.switch_mlp.up_proj": {
2766
+ "group_size": 32,
2767
+ "bits": 4,
2768
+ "mode": "mxfp4"
2769
+ },
2770
+ "model.layers.38.mlp.switch_mlp.down_proj": {
2771
+ "group_size": 32,
2772
+ "bits": 4,
2773
+ "mode": "mxfp4"
2774
+ },
2775
+ "model.layers.39.self_attn.q_proj": {
2776
+ "group_size": 32,
2777
+ "bits": 4,
2778
+ "mode": "mxfp4"
2779
+ },
2780
+ "model.layers.39.self_attn.k_proj": {
2781
+ "group_size": 32,
2782
+ "bits": 4,
2783
+ "mode": "mxfp4"
2784
+ },
2785
+ "model.layers.39.self_attn.v_proj": {
2786
+ "group_size": 32,
2787
+ "bits": 4,
2788
+ "mode": "mxfp4"
2789
+ },
2790
+ "model.layers.39.self_attn.o_proj": {
2791
+ "group_size": 32,
2792
+ "bits": 4,
2793
+ "mode": "mxfp4"
2794
+ },
2795
+ "model.layers.39.mlp.switch_mlp.gate_proj": {
2796
+ "group_size": 32,
2797
+ "bits": 4,
2798
+ "mode": "mxfp4"
2799
+ },
2800
+ "model.layers.39.mlp.switch_mlp.up_proj": {
2801
+ "group_size": 32,
2802
+ "bits": 4,
2803
+ "mode": "mxfp4"
2804
+ },
2805
+ "model.layers.39.mlp.switch_mlp.down_proj": {
2806
+ "group_size": 64,
2807
+ "bits": 6,
2808
+ "mode": "affine"
2809
+ },
2810
+ "model.layers.40.linear_attn.in_proj_qkvz": {
2811
+ "group_size": 32,
2812
+ "bits": 4,
2813
+ "mode": "mxfp4"
2814
+ },
2815
+ "model.layers.40.linear_attn.in_proj_ba": {
2816
+ "group_size": 64,
2817
+ "bits": 4,
2818
+ "mode": "affine"
2819
+ },
2820
+ "model.layers.40.linear_attn.out_proj": {
2821
+ "group_size": 64,
2822
+ "bits": 5,
2823
+ "mode": "affine"
2824
+ },
2825
+ "model.layers.40.mlp.switch_mlp.gate_proj": {
2826
+ "group_size": 32,
2827
+ "bits": 4,
2828
+ "mode": "mxfp4"
2829
+ },
2830
+ "model.layers.40.mlp.switch_mlp.up_proj": {
2831
+ "group_size": 32,
2832
+ "bits": 4,
2833
+ "mode": "mxfp4"
2834
+ },
2835
+ "model.layers.40.mlp.switch_mlp.down_proj": {
2836
+ "group_size": 32,
2837
+ "bits": 4,
2838
+ "mode": "mxfp4"
2839
+ },
2840
+ "model.layers.41.linear_attn.in_proj_qkvz": {
2841
+ "group_size": 32,
2842
+ "bits": 4,
2843
+ "mode": "mxfp4"
2844
+ },
2845
+ "model.layers.41.linear_attn.in_proj_ba": {
2846
+ "group_size": 64,
2847
+ "bits": 4,
2848
+ "mode": "affine"
2849
+ },
2850
+ "model.layers.41.linear_attn.out_proj": {
2851
+ "group_size": 64,
2852
+ "bits": 5,
2853
+ "mode": "affine"
2854
+ },
2855
+ "model.layers.41.mlp.switch_mlp.gate_proj": {
2856
+ "group_size": 32,
2857
+ "bits": 4,
2858
+ "mode": "mxfp4"
2859
+ },
2860
+ "model.layers.41.mlp.switch_mlp.up_proj": {
2861
+ "group_size": 32,
2862
+ "bits": 4,
2863
+ "mode": "mxfp4"
2864
+ },
2865
+ "model.layers.41.mlp.switch_mlp.down_proj": {
2866
+ "group_size": 32,
2867
+ "bits": 4,
2868
+ "mode": "mxfp4"
2869
+ },
2870
+ "model.layers.42.linear_attn.in_proj_qkvz": {
2871
+ "group_size": 32,
2872
+ "bits": 4,
2873
+ "mode": "mxfp4"
2874
+ },
2875
+ "model.layers.42.linear_attn.in_proj_ba": {
2876
+ "group_size": 64,
2877
+ "bits": 4,
2878
+ "mode": "affine"
2879
+ },
2880
+ "model.layers.42.linear_attn.out_proj": {
2881
+ "group_size": 64,
2882
+ "bits": 5,
2883
+ "mode": "affine"
2884
+ },
2885
+ "model.layers.42.mlp.switch_mlp.gate_proj": {
2886
+ "group_size": 32,
2887
+ "bits": 4,
2888
+ "mode": "mxfp4"
2889
+ },
2890
+ "model.layers.42.mlp.switch_mlp.up_proj": {
2891
+ "group_size": 32,
2892
+ "bits": 4,
2893
+ "mode": "mxfp4"
2894
+ },
2895
+ "model.layers.42.mlp.switch_mlp.down_proj": {
2896
+ "group_size": 64,
2897
+ "bits": 6,
2898
+ "mode": "affine"
2899
+ },
2900
+ "model.layers.43.self_attn.q_proj": {
2901
+ "group_size": 64,
2902
+ "bits": 5,
2903
+ "mode": "affine"
2904
+ },
2905
+ "model.layers.43.self_attn.k_proj": {
2906
+ "group_size": 64,
2907
+ "bits": 5,
2908
+ "mode": "affine"
2909
+ },
2910
+ "model.layers.43.self_attn.v_proj": {
2911
+ "group_size": 64,
2912
+ "bits": 5,
2913
+ "mode": "affine"
2914
+ },
2915
+ "model.layers.43.self_attn.o_proj": {
2916
+ "group_size": 64,
2917
+ "bits": 5,
2918
+ "mode": "affine"
2919
+ },
2920
+ "model.layers.43.mlp.switch_mlp.gate_proj": {
2921
+ "group_size": 32,
2922
+ "bits": 4,
2923
+ "mode": "mxfp4"
2924
+ },
2925
+ "model.layers.43.mlp.switch_mlp.up_proj": {
2926
+ "group_size": 32,
2927
+ "bits": 4,
2928
+ "mode": "mxfp4"
2929
+ },
2930
+ "model.layers.43.mlp.switch_mlp.down_proj": {
2931
+ "group_size": 64,
2932
+ "bits": 6,
2933
+ "mode": "affine"
2934
+ },
2935
+ "model.layers.44.linear_attn.in_proj_qkvz": {
2936
+ "group_size": 32,
2937
+ "bits": 4,
2938
+ "mode": "mxfp4"
2939
+ },
2940
+ "model.layers.44.linear_attn.in_proj_ba": {
2941
+ "group_size": 64,
2942
+ "bits": 4,
2943
+ "mode": "affine"
2944
+ },
2945
+ "model.layers.44.linear_attn.out_proj": {
2946
+ "group_size": 64,
2947
+ "bits": 6,
2948
+ "mode": "affine"
2949
+ },
2950
+ "model.layers.44.mlp.switch_mlp.gate_proj": {
2951
+ "group_size": 32,
2952
+ "bits": 4,
2953
+ "mode": "mxfp4"
2954
+ },
2955
+ "model.layers.44.mlp.switch_mlp.up_proj": {
2956
+ "group_size": 32,
2957
+ "bits": 4,
2958
+ "mode": "mxfp4"
2959
+ },
2960
+ "model.layers.44.mlp.switch_mlp.down_proj": {
2961
+ "group_size": 32,
2962
+ "bits": 4,
2963
+ "mode": "mxfp4"
2964
+ },
2965
+ "model.layers.45.linear_attn.in_proj_qkvz": {
2966
+ "group_size": 64,
2967
+ "bits": 5,
2968
+ "mode": "affine"
2969
+ },
2970
+ "model.layers.45.linear_attn.in_proj_ba": {
2971
+ "group_size": 64,
2972
+ "bits": 4,
2973
+ "mode": "affine"
2974
+ },
2975
+ "model.layers.45.linear_attn.out_proj": {
2976
+ "group_size": 64,
2977
+ "bits": 8,
2978
+ "mode": "affine"
2979
+ },
2980
+ "model.layers.45.mlp.switch_mlp.gate_proj": {
2981
+ "group_size": 32,
2982
+ "bits": 4,
2983
+ "mode": "mxfp4"
2984
+ },
2985
+ "model.layers.45.mlp.switch_mlp.up_proj": {
2986
+ "group_size": 32,
2987
+ "bits": 4,
2988
+ "mode": "mxfp4"
2989
+ },
2990
+ "model.layers.45.mlp.switch_mlp.down_proj": {
2991
+ "group_size": 64,
2992
+ "bits": 6,
2993
+ "mode": "affine"
2994
+ },
2995
+ "model.layers.46.linear_attn.in_proj_qkvz": {
2996
+ "group_size": 64,
2997
+ "bits": 5,
2998
+ "mode": "affine"
2999
+ },
3000
+ "model.layers.46.linear_attn.in_proj_ba": {
3001
+ "group_size": 64,
3002
+ "bits": 4,
3003
+ "mode": "affine"
3004
+ },
3005
+ "model.layers.46.linear_attn.out_proj": {
3006
+ "group_size": 64,
3007
+ "bits": 8,
3008
+ "mode": "affine"
3009
+ },
3010
+ "model.layers.46.mlp.switch_mlp.gate_proj": {
3011
+ "group_size": 32,
3012
+ "bits": 4,
3013
+ "mode": "mxfp4"
3014
+ },
3015
+ "model.layers.46.mlp.switch_mlp.up_proj": {
3016
+ "group_size": 32,
3017
+ "bits": 4,
3018
+ "mode": "mxfp4"
3019
+ },
3020
+ "model.layers.46.mlp.switch_mlp.down_proj": {
3021
+ "group_size": 64,
3022
+ "bits": 6,
3023
+ "mode": "affine"
3024
+ },
3025
+ "model.layers.47.self_attn.q_proj": {
3026
+ "group_size": 64,
3027
+ "bits": 5,
3028
+ "mode": "affine"
3029
+ },
3030
+ "model.layers.47.self_attn.k_proj": {
3031
+ "group_size": 64,
3032
+ "bits": 5,
3033
+ "mode": "affine"
3034
+ },
3035
+ "model.layers.47.self_attn.v_proj": {
3036
+ "group_size": 64,
3037
+ "bits": 5,
3038
+ "mode": "affine"
3039
+ },
3040
+ "model.layers.47.self_attn.o_proj": {
3041
+ "group_size": 64,
3042
+ "bits": 5,
3043
+ "mode": "affine"
3044
+ },
3045
+ "model.layers.47.mlp.switch_mlp.gate_proj": {
3046
+ "group_size": 32,
3047
+ "bits": 4,
3048
+ "mode": "mxfp4"
3049
+ },
3050
+ "model.layers.47.mlp.switch_mlp.up_proj": {
3051
+ "group_size": 32,
3052
+ "bits": 4,
3053
+ "mode": "mxfp4"
3054
+ },
3055
+ "model.layers.47.mlp.switch_mlp.down_proj": {
3056
+ "group_size": 64,
3057
+ "bits": 6,
3058
+ "mode": "affine"
3059
+ },
3060
+ "lm_head": {
3061
+ "group_size": 64,
3062
+ "bits": 6,
3063
+ "mode": "affine"
3064
+ }
3065
+ },
3066
+ "rms_norm_eps": 1e-06,
3067
+ "rope_scaling": null,
3068
+ "rope_theta": 5000000,
3069
+ "router_aux_loss_coef": 0.001,
3070
+ "shared_expert_intermediate_size": 512,
3071
+ "tie_word_embeddings": false,
3072
+ "torch_dtype": "bfloat16",
3073
+ "transformers_version": "4.57.0.dev0",
3074
+ "use_cache": true,
3075
+ "use_sliding_window": false,
3076
+ "vocab_size": 151936
3077
+ }
generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token_id": 151643,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 151645,
6
+ 151643
7
+ ],
8
+ "pad_token_id": 151643,
9
+ "temperature": 1.0,
10
+ "top_k": 40,
11
+ "top_p": 0.95,
12
+ "transformers_version": "4.57.3"
13
+ }
model-00001-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2affc16c11733e8eb69ff481afbf1b96fcc2feae90fdb93aec4aa7fc05c165b9
3
+ size 5307013181
model-00002-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d9f1cd72bf80e0aa83409d81969ee4350a75a53d8713a58a13e764ce2a8feede
3
+ size 5279957955
model-00003-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9387f3eb3deafc76d59c96c5023072b10a1c7800853bec8ddc59e14ee51d3f3a
3
+ size 5136900516
model-00004-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:668b6884c6523d380486e336abb1f0bd1047fa122afa6670e331b41f6eeb9fe8
3
+ size 5362257844
model-00005-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:13a69d9d11475ff9755f2dde7aedded137d22f93565f37dbb9b44367be242f19
3
+ size 5319890802
model-00006-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16ebb1a6599085e807e6d769df86d7eb7a7c18168eda40c75f724a8e14c7aad1
3
+ size 5113745299
model-00007-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61d131f5b3f141d9fa9d2fe9f0a79ca3c570f943143b7c295c6293f77ae073c1
3
+ size 5128962934
model-00008-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:730e135811db136ec15460c1e9cfeb80661736be2f54ca05d94b584601d17782
3
+ size 4988768634
model-00009-of-00009.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b3aa283b96888a44073b2aeb162d28f8d73f246c8af067a7af1de3dd16b806d6
3
+ size 2756930749
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
qwen3_coder_detector_sgl.py ADDED
@@ -0,0 +1,474 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ast
2
+ import json
3
+ import logging
4
+ import re
5
+ from typing import Any, List, Optional
6
+
7
+ from sglang.srt.entrypoints.openai.protocol import Tool
8
+ from sglang.srt.function_call.base_format_detector import BaseFormatDetector
9
+ from sglang.srt.function_call.core_types import (
10
+ StreamingParseResult,
11
+ ToolCallItem,
12
+ _GetInfoFunc,
13
+ )
14
+
15
+ logger = logging.getLogger(__name__)
16
+
17
+
18
+ class Qwen3CoderDetector(BaseFormatDetector):
19
+ def __init__(self):
20
+ super().__init__()
21
+
22
+ # Sentinel tokens
23
+ self.tool_call_start_token: str = "<tool_call>"
24
+ self.tool_call_end_token: str = "</tool_call>"
25
+ self.tool_call_prefix: str = "<function="
26
+ self.function_end_token: str = "</function>"
27
+ self.parameter_prefix: str = "<parameter="
28
+ self.parameter_end_token: str = "</parameter>"
29
+
30
+ # Regex for non-streaming fallback
31
+ self.tool_call_regex = re.compile(r"<tool_call>(.*?)</tool_call>", re.DOTALL)
32
+ self.tool_call_function_regex = re.compile(
33
+ r"<function=(.*?)</function>|<function=(.*)$", re.DOTALL
34
+ )
35
+ self.tool_call_parameter_regex = re.compile(
36
+ r"<parameter=(.*?)(?:</parameter>|(?=<parameter=)|(?=</function>)|$)",
37
+ re.DOTALL,
38
+ )
39
+
40
+ # Streaming State
41
+ # Base class already initializes _buffer, we just use it directly
42
+ # No need to check with hasattr - we control the lifecycle through inheritance
43
+
44
+ # Index pointing to the next character to be processed in buffer
45
+ self.parsed_pos: int = 0
46
+ # Parameter count inside the current tool being processed, used to determine whether to add comma
47
+ self.current_tool_param_count: int = 0
48
+ # Flag indicating whether current tool has already sent '{'
49
+ self.json_started: bool = False
50
+
51
+ # [FIX] New state flag: mark whether inside tool_call structure block
52
+ self.is_inside_tool_call: bool = False
53
+
54
+ # Initialize attributes that were missing in the original PR
55
+ self.current_func_name: Optional[str] = None
56
+
57
+ def has_tool_call(self, text: str) -> bool:
58
+ return self.tool_call_start_token in text
59
+
60
+ def _get_arguments_config(
61
+ self, func_name: str, tools: Optional[list[Tool]]
62
+ ) -> dict:
63
+ """Extract argument configuration for a function."""
64
+ if tools is None:
65
+ return {}
66
+ for config in tools:
67
+ try:
68
+ config_type = config.type
69
+ config_function = config.function
70
+ config_function_name = config_function.name
71
+ except AttributeError:
72
+ continue
73
+
74
+ if config_type == "function" and config_function_name == func_name:
75
+ try:
76
+ params = config_function.parameters
77
+ except AttributeError:
78
+ return {}
79
+
80
+ if isinstance(params, dict) and "properties" in params:
81
+ return params["properties"]
82
+ elif isinstance(params, dict):
83
+ return params
84
+ else:
85
+ return {}
86
+ logger.warning(f"Tool '{func_name}' is not defined in the tools list.")
87
+ return {}
88
+
89
+ def _convert_param_value(
90
+ self, param_value: str, param_name: str, param_config: dict, func_name: str
91
+ ) -> Any:
92
+ """Convert parameter value based on its type in the schema."""
93
+ # Handle null value for any type
94
+ if param_value.lower() == "null":
95
+ return None
96
+
97
+ if param_name not in param_config:
98
+ if param_config != {}:
99
+ logger.warning(
100
+ f"Parsed parameter '{param_name}' is not defined in the tool "
101
+ f"parameters for tool '{func_name}', directly returning the string value."
102
+ )
103
+ return param_value
104
+
105
+ if (
106
+ isinstance(param_config[param_name], dict)
107
+ and "type" in param_config[param_name]
108
+ ):
109
+ param_type = str(param_config[param_name]["type"]).strip().lower()
110
+ else:
111
+ param_type = "string"
112
+ if param_type in ["string", "str", "text", "varchar", "char", "enum"]:
113
+ return param_value
114
+ elif (
115
+ param_type.startswith("int")
116
+ or param_type.startswith("uint")
117
+ or param_type.startswith("long")
118
+ or param_type.startswith("short")
119
+ or param_type.startswith("unsigned")
120
+ ):
121
+ try:
122
+ param_value = int(param_value)
123
+ except Exception:
124
+ logger.warning(
125
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not an integer in tool "
126
+ f"'{func_name}', degenerating to string."
127
+ )
128
+ return param_value
129
+ elif param_type.startswith("num") or param_type.startswith("float"):
130
+ try:
131
+ maybe_convert = (
132
+ False if "." in param_value or "e" in param_value.lower() else True
133
+ )
134
+ param_value: float = float(param_value)
135
+ if maybe_convert and param_value.is_integer():
136
+ param_value = int(param_value)
137
+ except Exception:
138
+ logger.warning(
139
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not a float in tool "
140
+ f"'{func_name}', degenerating to string."
141
+ )
142
+ return param_value
143
+ elif param_type in ["boolean", "bool", "binary"]:
144
+ param_value = param_value.lower()
145
+ if param_value not in ["true", "false"]:
146
+ logger.warning(
147
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not a boolean (`true` of `false`) in tool '{func_name}', degenerating to false."
148
+ )
149
+ return param_value == "true"
150
+ else:
151
+ if (
152
+ param_type in ["object", "array", "arr"]
153
+ or param_type.startswith("dict")
154
+ or param_type.startswith("list")
155
+ ):
156
+ try:
157
+ param_value = json.loads(param_value)
158
+ return param_value
159
+ except Exception:
160
+ logger.warning(
161
+ f"Parsed value '{param_value}' of parameter '{param_name}' cannot be parsed with json.loads in tool "
162
+ f"'{func_name}', will try other methods to parse it."
163
+ )
164
+ try:
165
+ param_value = ast.literal_eval(param_value) # safer
166
+ except Exception:
167
+ logger.warning(
168
+ f"Parsed value '{param_value}' of parameter '{param_name}' cannot be converted via Python `ast.literal_eval()` in tool '{func_name}', degenerating to string."
169
+ )
170
+ return param_value
171
+
172
+ def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult:
173
+ """One-shot parsing for non-streaming scenarios."""
174
+ if self.tool_call_start_token not in text:
175
+ return StreamingParseResult(normal_text=text)
176
+
177
+ calls = []
178
+ try:
179
+ # Simple cleanup of the text to find tool calls
180
+ # Note: This is a simplified regex approach consistent with vLLM
181
+ raw_tool_calls = self.tool_call_regex.findall(text)
182
+ if not raw_tool_calls:
183
+ # Fallback: maybe the whole text is inside the tag or tags are stripped
184
+ if self.tool_call_prefix in text:
185
+ raw_tool_calls = [text]
186
+
187
+ tool_idx = 0
188
+ for tool_content in raw_tool_calls:
189
+ # Find function calls
190
+ funcs = self.tool_call_function_regex.findall(tool_content)
191
+ for func_match in funcs:
192
+ func_body = func_match[0] or func_match[1]
193
+ if ">" not in func_body:
194
+ continue
195
+
196
+ name_end = func_body.index(">")
197
+ func_name = func_body[:name_end]
198
+ params_str = func_body[name_end + 1 :]
199
+
200
+ param_config = self._get_arguments_config(func_name, tools)
201
+ parsed_params = {}
202
+
203
+ for p_match in self.tool_call_parameter_regex.findall(params_str):
204
+ if ">" not in p_match:
205
+ continue
206
+ p_idx = p_match.index(">")
207
+ p_name = p_match[:p_idx]
208
+ p_val = p_match[p_idx + 1 :]
209
+ # Remove prefixing and trailing \n
210
+ if p_val.startswith("\n"):
211
+ p_val = p_val[1:]
212
+ if p_val.endswith("\n"):
213
+ p_val = p_val[:-1]
214
+
215
+ parsed_params[p_name] = self._convert_param_value(
216
+ p_val, p_name, param_config, func_name
217
+ )
218
+
219
+ calls.append(
220
+ ToolCallItem(
221
+ tool_index=tool_idx,
222
+ name=func_name,
223
+ parameters=json.dumps(parsed_params, ensure_ascii=False),
224
+ )
225
+ )
226
+ tool_idx += 1
227
+
228
+ # Determine normal text (text before the first tool call)
229
+ start_idx = text.find(self.tool_call_start_token)
230
+ if start_idx == -1:
231
+ start_idx = text.find(self.tool_call_prefix)
232
+ normal_text = text[:start_idx] if start_idx > 0 else ""
233
+
234
+ return StreamingParseResult(normal_text=normal_text, calls=calls)
235
+
236
+ except Exception as e:
237
+ logger.error(f"Error in detect_and_parse: {e}")
238
+ return StreamingParseResult(normal_text=text)
239
+
240
+ def parse_streaming_increment(
241
+ self, new_text: str, tools: List[Tool]
242
+ ) -> StreamingParseResult:
243
+ """
244
+ Robust cursor-based streaming parser.
245
+ """
246
+ self._buffer += new_text
247
+
248
+ # Guard against empty buffer
249
+ if not self._buffer:
250
+ return StreamingParseResult()
251
+
252
+ calls = []
253
+ normal_text_chunks = []
254
+
255
+ while True:
256
+ # Working text slice
257
+ current_slice = self._buffer[self.parsed_pos :]
258
+
259
+ # Optimization: If almost empty, wait for more
260
+ if not current_slice:
261
+ break
262
+
263
+ # -------------------------------------------------------
264
+ # 1. Priority detection: check if it's the start of Tool Call
265
+ # -------------------------------------------------------
266
+ if current_slice.startswith(self.tool_call_start_token):
267
+ self.parsed_pos += len(self.tool_call_start_token)
268
+ self.is_inside_tool_call = True
269
+ continue
270
+
271
+ # -------------------------------------------------------
272
+ # 2. Function Name: <function=name>
273
+ # -------------------------------------------------------
274
+ if current_slice.startswith(self.tool_call_prefix):
275
+ end_angle = current_slice.find(">")
276
+ if end_angle != -1:
277
+ func_name = current_slice[len(self.tool_call_prefix) : end_angle]
278
+
279
+ self.current_tool_id += 1
280
+ self.current_tool_name_sent = True
281
+ self.current_tool_param_count = 0
282
+ self.json_started = False
283
+ self.current_func_name = func_name
284
+
285
+ calls.append(
286
+ ToolCallItem(
287
+ tool_index=self.current_tool_id,
288
+ name=func_name,
289
+ parameters="",
290
+ )
291
+ )
292
+
293
+ self.parsed_pos += end_angle + 1
294
+ continue
295
+ else:
296
+ # Incomplete tag
297
+ break
298
+
299
+ # -------------------------------------------------------
300
+ # 3. Parameter: <parameter=name>value...
301
+ # -------------------------------------------------------
302
+ if current_slice.startswith(self.parameter_prefix):
303
+ name_end = current_slice.find(">")
304
+ if name_end != -1:
305
+ value_start_idx = name_end + 1
306
+ rest_of_slice = current_slice[value_start_idx:]
307
+
308
+ # A parameter can end in multiple ways:
309
+ # 1. [Normal] Encounter </parameter>
310
+ # 2. [Abnormal] Encounter next <parameter=
311
+ # 3. [Abnormal] Encounter </function>
312
+ # So we need to find the smallest one as the parameter end position.
313
+ cand_end_param = rest_of_slice.find(self.parameter_end_token)
314
+ cand_next_param = rest_of_slice.find(self.parameter_prefix)
315
+ cand_end_func = rest_of_slice.find(self.function_end_token)
316
+
317
+ candidates = []
318
+ if cand_end_param != -1:
319
+ candidates.append(
320
+ (cand_end_param, len(self.parameter_end_token))
321
+ )
322
+ if cand_next_param != -1:
323
+ candidates.append((cand_next_param, 0))
324
+ if cand_end_func != -1:
325
+ candidates.append((cand_end_func, 0))
326
+
327
+ if candidates:
328
+ best_cand = min(candidates, key=lambda x: x[0])
329
+ end_pos = best_cand[0]
330
+ end_token_len = best_cand[1]
331
+
332
+ param_name = current_slice[
333
+ len(self.parameter_prefix) : name_end
334
+ ]
335
+ raw_value = rest_of_slice[:end_pos]
336
+
337
+ # Cleanup value
338
+ if raw_value.startswith("\n"):
339
+ raw_value = raw_value[1:]
340
+ if raw_value.endswith("\n"):
341
+ raw_value = raw_value[:-1]
342
+
343
+ # JSON Construction
344
+ if not self.json_started:
345
+ calls.append(
346
+ ToolCallItem(
347
+ tool_index=self.current_tool_id, parameters="{"
348
+ )
349
+ )
350
+ self.json_started = True
351
+
352
+ param_config = self._get_arguments_config(
353
+ self.current_func_name, tools
354
+ )
355
+ converted_val = self._convert_param_value(
356
+ raw_value, param_name, param_config, self.current_func_name
357
+ )
358
+
359
+ # Construct JSON fragment: "key": value
360
+ # Note: We must be careful with json.dumps to ensure valid JSON streaming
361
+ json_key_val = f"{json.dumps(param_name)}: {json.dumps(converted_val, ensure_ascii=False)}"
362
+
363
+ if self.current_tool_param_count > 0:
364
+ fragment = f", {json_key_val}"
365
+ else:
366
+ fragment = json_key_val
367
+
368
+ calls.append(
369
+ ToolCallItem(
370
+ tool_index=self.current_tool_id, parameters=fragment
371
+ )
372
+ )
373
+ self.current_tool_param_count += 1
374
+
375
+ # Advance cursor
376
+ total_len = (name_end + 1) + end_pos + end_token_len
377
+ self.parsed_pos += total_len
378
+ continue
379
+
380
+ # Incomplete parameter tag or value
381
+ break
382
+
383
+ # -------------------------------------------------------
384
+ # 4. Function End: </function>
385
+ # -------------------------------------------------------
386
+ if current_slice.startswith(self.function_end_token):
387
+ if not self.json_started:
388
+ calls.append(
389
+ ToolCallItem(tool_index=self.current_tool_id, parameters="{")
390
+ )
391
+ self.json_started = True
392
+
393
+ calls.append(
394
+ ToolCallItem(tool_index=self.current_tool_id, parameters="}")
395
+ )
396
+ self.parsed_pos += len(self.function_end_token)
397
+ self.current_func_name = None
398
+ continue
399
+
400
+ # -------------------------------------------------------
401
+ # 5. Tool Call End: </tool_call>
402
+ # -------------------------------------------------------
403
+ if current_slice.startswith(self.tool_call_end_token):
404
+ self.parsed_pos += len(self.tool_call_end_token)
405
+ self.is_inside_tool_call = False # [FIX] Exit tool call region
406
+ continue
407
+
408
+ # -------------------------------------------------------
409
+ # 6. Handling content / whitespace / normal text
410
+ # -------------------------------------------------------
411
+ # If current position is not the start of a tag (i.e., doesn't start with <), it might be plain text,
412
+ # or a newline between two tags.
413
+ # But we need to be careful not to output truncated tags like "<fun" as text.
414
+
415
+ next_open_angle = current_slice.find("<")
416
+
417
+ if next_open_angle == -1:
418
+ # This entire segment is plain text
419
+ if not self.is_inside_tool_call:
420
+ normal_text_chunks.append(current_slice)
421
+ # [FIX] If inside tool call, discard this text (usually \n), don't append
422
+ self.parsed_pos += len(current_slice)
423
+ continue
424
+
425
+ elif next_open_angle == 0:
426
+ # Looks like a Tag, but doesn't match any known Tag above
427
+
428
+ possible_tags = [
429
+ self.tool_call_start_token,
430
+ self.tool_call_end_token,
431
+ self.tool_call_prefix,
432
+ self.function_end_token,
433
+ self.parameter_prefix,
434
+ self.parameter_end_token,
435
+ ]
436
+
437
+ is_potential_tag = False
438
+ for tag in possible_tags:
439
+ if tag.startswith(current_slice):
440
+ is_potential_tag = True
441
+ break
442
+
443
+ if is_potential_tag:
444
+ break # Wait for more
445
+ else:
446
+ # Just a plain '<' symbol
447
+ if not self.is_inside_tool_call:
448
+ normal_text_chunks.append("<")
449
+ self.parsed_pos += 1
450
+ continue
451
+
452
+ else:
453
+ # '<' is in the middle
454
+ text_segment = current_slice[:next_open_angle]
455
+ if not self.is_inside_tool_call:
456
+ normal_text_chunks.append(text_segment)
457
+ # [FIX] If inside tool call, discard whitespace/text before Tag
458
+ self.parsed_pos += next_open_angle
459
+ continue
460
+
461
+ # Memory Cleanup: Slice the buffer
462
+ # Keep unparsed part, discard parsed part
463
+ if self.parsed_pos > 0:
464
+ self._buffer = self._buffer[self.parsed_pos :]
465
+ self.parsed_pos = 0
466
+
467
+ normal_text = "".join(normal_text_chunks) if normal_text_chunks else ""
468
+ return StreamingParseResult(calls=calls, normal_text=normal_text)
469
+
470
+ def supports_structural_tag(self) -> bool:
471
+ return False
472
+
473
+ def structure_info(self) -> _GetInfoFunc:
474
+ raise NotImplementedError
qwen3coder_tool_parser_vllm.py ADDED
@@ -0,0 +1,690 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # SPDX-License-Identifier: Apache-2.0
2
+ # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+ import ast
4
+ import json
5
+ import uuid
6
+ from collections.abc import Sequence
7
+ from typing import Any, List, Optional, Union
8
+
9
+ import regex as re
10
+
11
+ from vllm.entrypoints.openai.protocol import (ChatCompletionRequest,
12
+ ChatCompletionToolsParam,
13
+ DeltaFunctionCall, DeltaMessage,
14
+ DeltaToolCall,
15
+ ExtractedToolCallInformation,
16
+ FunctionCall, ToolCall)
17
+ from vllm.entrypoints.openai.tool_parsers.abstract_tool_parser import (
18
+ ToolParser, ToolParserManager)
19
+ from vllm.logger import init_logger
20
+ from vllm.transformers_utils.tokenizer import AnyTokenizer
21
+
22
+ logger = init_logger(__name__)
23
+
24
+
25
+ @ToolParserManager.register_module("qwen3_coder")
26
+ class Qwen3CoderToolParser(ToolParser):
27
+
28
+ def __init__(self, tokenizer: AnyTokenizer):
29
+ super().__init__(tokenizer)
30
+
31
+ self.current_tool_name_sent: bool = False
32
+ self.prev_tool_call_arr: list[dict] = []
33
+ self.current_tool_id: int = -1
34
+ self.streamed_args_for_tool: list[str] = []
35
+
36
+ # Sentinel tokens for streaming mode
37
+ self.tool_call_start_token: str = "<tool_call>"
38
+ self.tool_call_end_token: str = "</tool_call>"
39
+ self.tool_call_prefix: str = "<function="
40
+ self.function_end_token: str = "</function>"
41
+ self.parameter_prefix: str = "<parameter="
42
+ self.parameter_end_token: str = "</parameter>"
43
+ self.is_tool_call_started: bool = False
44
+ self.failed_count: int = 0
45
+
46
+ # Enhanced streaming state - reset for each new message
47
+ self._reset_streaming_state()
48
+
49
+ # Regex patterns
50
+ self.tool_call_complete_regex = re.compile(
51
+ r"<tool_call>(.*?)</tool_call>", re.DOTALL)
52
+ self.tool_call_regex = re.compile(
53
+ r"<tool_call>(.*?)</tool_call>|<tool_call>(.*?)$", re.DOTALL)
54
+ self.tool_call_function_regex = re.compile(
55
+ r"<function=(.*?)</function>|<function=(.*)$", re.DOTALL)
56
+ self.tool_call_parameter_regex = re.compile(
57
+ r"<parameter=(.*?)(?:</parameter>|(?=<parameter=)|(?=</function>)|$)",
58
+ re.DOTALL)
59
+
60
+ if not self.model_tokenizer:
61
+ raise ValueError(
62
+ "The model tokenizer must be passed to the ToolParser "
63
+ "constructor during construction.")
64
+
65
+ self.tool_call_start_token_id = self.vocab.get(
66
+ self.tool_call_start_token)
67
+ self.tool_call_end_token_id = self.vocab.get(self.tool_call_end_token)
68
+
69
+ if self.tool_call_start_token_id is None or self.tool_call_end_token_id is None:
70
+ raise RuntimeError(
71
+ "Qwen3 XML Tool parser could not locate tool call start/end "
72
+ "tokens in the tokenizer!")
73
+
74
+ logger.info(
75
+ f"vLLM Successfully import tool parser {self.__class__.__name__} !"
76
+ )
77
+
78
+ def _generate_tool_call_id(self) -> str:
79
+ """Generate a unique tool call ID."""
80
+ return f"call_{uuid.uuid4().hex[:24]}"
81
+
82
+ def _reset_streaming_state(self):
83
+ """Reset all streaming state."""
84
+ self.current_tool_index = 0
85
+ self.is_tool_call_started = False
86
+ self.header_sent = False
87
+ self.current_tool_id = None
88
+ self.current_function_name = None
89
+ self.current_param_name = None
90
+ self.current_param_value = ""
91
+ self.param_count = 0
92
+ self.in_param = False
93
+ self.in_function = False
94
+ self.accumulated_text = ""
95
+ self.json_started = False
96
+ self.json_closed = False
97
+ # Store accumulated parameters for type conversion
98
+ self.accumulated_params = {}
99
+ self.streaming_request = None
100
+
101
+ def _get_arguments_config(
102
+ self, func_name: str,
103
+ tools: Optional[list[ChatCompletionToolsParam]]) -> dict:
104
+ """Extract argument configuration for a function."""
105
+ if tools is None:
106
+ return {}
107
+ for config in tools:
108
+ if not hasattr(config, "type") or not (hasattr(
109
+ config, "function") and hasattr(config.function, "name")):
110
+ continue
111
+ if config.type == "function" and config.function.name == func_name:
112
+ if not hasattr(config.function, "parameters"):
113
+ return {}
114
+ params = config.function.parameters
115
+ if isinstance(params, dict) and "properties" in params:
116
+ return params["properties"]
117
+ elif isinstance(params, dict):
118
+ return params
119
+ else:
120
+ return {}
121
+ logger.warning(f"Tool '{func_name}' is not defined in the tools list.")
122
+ return {}
123
+
124
+ def _convert_param_value(self, param_value: str, param_name: str,
125
+ param_config: dict, func_name: str) -> Any:
126
+ """Convert parameter value based on its type in the schema."""
127
+ # Handle null value for any type
128
+ if param_value.lower() == "null":
129
+ return None
130
+
131
+ if param_name not in param_config:
132
+ if param_config != {}:
133
+ logger.warning(
134
+ f"Parsed parameter '{param_name}' is not defined in the tool "
135
+ f"parameters for tool '{func_name}', directly returning the string value."
136
+ )
137
+ return param_value
138
+
139
+ if isinstance(param_config[param_name],
140
+ dict) and "type" in param_config[param_name]:
141
+ param_type = str(param_config[param_name]["type"]).strip().lower()
142
+ else:
143
+ param_type = "string"
144
+ if param_type in ["string", "str", "text", "varchar", "char", "enum"]:
145
+ return param_value
146
+ elif param_type.startswith("int") or param_type.startswith(
147
+ "uint") or param_type.startswith(
148
+ "long") or param_type.startswith(
149
+ "short") or param_type.startswith("unsigned"):
150
+ try:
151
+ param_value = int(param_value)
152
+ except:
153
+ logger.warning(
154
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not an integer in tool "
155
+ f"'{func_name}', degenerating to string.")
156
+ return param_value
157
+ elif param_type.startswith("num") or param_type.startswith("float"):
158
+ try:
159
+ maybe_convert = False if "." in param_value or "e" in param_value.lower() else True
160
+ param_value: float = float(param_value)
161
+ if maybe_convert and param_value.is_integer():
162
+ param_value = int(param_value)
163
+ except:
164
+ logger.warning(
165
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not a float in tool "
166
+ f"'{func_name}', degenerating to string.")
167
+ return param_value
168
+ elif param_type in ["boolean", "bool", "binary"]:
169
+ param_value = param_value.lower()
170
+ if param_value not in ["true", "false"]:
171
+ logger.warning(
172
+ f"Parsed value '{param_value}' of parameter '{param_name}' is not a boolean (`true` of `false`) in tool '{func_name}', degenerating to false."
173
+ )
174
+ return param_value == "true"
175
+ else:
176
+ if param_type in ["object", "array", "arr"
177
+ ] or param_type.startswith(
178
+ "dict") or param_type.startswith("list"):
179
+ try:
180
+ param_value = json.loads(param_value)
181
+ return param_value
182
+ except:
183
+ logger.warning(
184
+ f"Parsed value '{param_value}' of parameter '{param_name}' cannot be parsed with json.loads in tool "
185
+ f"'{func_name}', will try other methods to parse it.")
186
+ try:
187
+ param_value = ast.literal_eval(param_value) # safer
188
+ except:
189
+ logger.warning(
190
+ f"Parsed value '{param_value}' of parameter '{param_name}' cannot be converted via Python `ast.literal_eval()` in tool '{func_name}', degenerating to string."
191
+ )
192
+ return param_value
193
+
194
+ def _parse_xml_function_call(
195
+ self, function_call_str: str,
196
+ tools: Optional[list[ChatCompletionToolsParam]]
197
+ ) -> Optional[ToolCall]:
198
+
199
+ # Extract function name
200
+ end_index = function_call_str.index(">")
201
+ function_name = function_call_str[:end_index]
202
+ param_config = self._get_arguments_config(function_name, tools)
203
+ parameters = function_call_str[end_index + 1:]
204
+ param_dict = {}
205
+ for match_text in self.tool_call_parameter_regex.findall(parameters):
206
+ idx = match_text.index(">")
207
+ param_name = match_text[:idx]
208
+ param_value = str(match_text[idx + 1:])
209
+ # Remove prefix and trailing \n
210
+ if param_value.startswith("\n"):
211
+ param_value = param_value[1:]
212
+ if param_value.endswith("\n"):
213
+ param_value = param_value[:-1]
214
+
215
+ param_dict[param_name] = self._convert_param_value(
216
+ param_value, param_name, param_config, function_name)
217
+ return ToolCall(
218
+ type="function",
219
+ function=FunctionCall(name=function_name,
220
+ arguments=json.dumps(param_dict,
221
+ ensure_ascii=False)),
222
+ )
223
+
224
+ def _get_function_calls(self, model_output: str) -> List[str]:
225
+ # Find all tool calls
226
+ matched_ranges = self.tool_call_regex.findall(model_output)
227
+ raw_tool_calls = [
228
+ match[0] if match[0] else match[1] for match in matched_ranges
229
+ ]
230
+
231
+ # Back-off strategy if no tool_call tags found
232
+ if len(raw_tool_calls) == 0:
233
+ raw_tool_calls = [model_output]
234
+
235
+ raw_function_calls = []
236
+ for tool_call in raw_tool_calls:
237
+ raw_function_calls.extend(
238
+ self.tool_call_function_regex.findall(tool_call))
239
+
240
+ function_calls = [
241
+ match[0] if match[0] else match[1] for match in raw_function_calls
242
+ ]
243
+ return function_calls
244
+
245
+ def extract_tool_calls(
246
+ self,
247
+ model_output: str,
248
+ request: ChatCompletionRequest,
249
+ ) -> ExtractedToolCallInformation:
250
+ # Quick check to avoid unnecessary processing
251
+ if self.tool_call_prefix not in model_output:
252
+ return ExtractedToolCallInformation(tools_called=False,
253
+ tool_calls=[],
254
+ content=model_output)
255
+
256
+ try:
257
+ function_calls = self._get_function_calls(model_output)
258
+ if len(function_calls) == 0:
259
+ return ExtractedToolCallInformation(tools_called=False,
260
+ tool_calls=[],
261
+ content=model_output)
262
+
263
+ tool_calls = [
264
+ self._parse_xml_function_call(function_call_str, request.tools)
265
+ for function_call_str in function_calls
266
+ ]
267
+
268
+ # Populate prev_tool_call_arr for serving layer to set finish_reason
269
+ self.prev_tool_call_arr.clear() # Clear previous calls
270
+ for tool_call in tool_calls:
271
+ if tool_call:
272
+ self.prev_tool_call_arr.append({
273
+ "name":
274
+ tool_call.function.name,
275
+ "arguments":
276
+ tool_call.function.arguments,
277
+ })
278
+
279
+ # Extract content before tool calls
280
+ content_index = model_output.find(self.tool_call_start_token)
281
+ content_index = content_index if content_index >= 0 else model_output.find(
282
+ self.tool_call_prefix)
283
+ content = model_output[:content_index] # .rstrip()
284
+
285
+ return ExtractedToolCallInformation(
286
+ tools_called=(len(tool_calls) > 0),
287
+ tool_calls=tool_calls,
288
+ content=content if content else None,
289
+ )
290
+
291
+ except Exception:
292
+ logger.exception("Error in extracting tool call from response.")
293
+ return ExtractedToolCallInformation(tools_called=False,
294
+ tool_calls=[],
295
+ content=model_output)
296
+
297
+ def extract_tool_calls_streaming(
298
+ self,
299
+ previous_text: str,
300
+ current_text: str,
301
+ delta_text: str,
302
+ previous_token_ids: Sequence[int],
303
+ current_token_ids: Sequence[int],
304
+ delta_token_ids: Sequence[int],
305
+ request: ChatCompletionRequest,
306
+ ) -> Union[DeltaMessage, None]:
307
+ # Store request for type conversion
308
+ if not previous_text:
309
+ self._reset_streaming_state()
310
+ self.streaming_request = request
311
+
312
+ # If no delta text, return None unless it's an EOS token after tool calls
313
+ if not delta_text:
314
+ # Check if this is an EOS token after all tool calls are complete
315
+ # We check for tool calls in the text even if is_tool_call_started is False
316
+ # because it might have been reset after processing all tools
317
+ if delta_token_ids and self.tool_call_end_token_id not in delta_token_ids:
318
+ # Count complete tool calls
319
+ complete_calls = len(
320
+ self.tool_call_complete_regex.findall(current_text))
321
+
322
+ # If we have completed tool calls and populated prev_tool_call_arr
323
+ if complete_calls > 0 and len(self.prev_tool_call_arr) > 0:
324
+ # Check if all tool calls are closed
325
+ open_calls = current_text.count(
326
+ self.tool_call_start_token) - current_text.count(
327
+ self.tool_call_end_token)
328
+ if open_calls == 0:
329
+ # Return empty delta message to allow finish_reason processing
330
+ return DeltaMessage(content="")
331
+ elif not self.is_tool_call_started and current_text:
332
+ # This is a regular content response that's now complete
333
+ return DeltaMessage(content="")
334
+ return None
335
+
336
+ # Update accumulated text
337
+ self.accumulated_text = current_text
338
+
339
+ # Check if we need to advance to next tool
340
+ if self.json_closed and not self.in_function:
341
+ # Check if this tool call has ended
342
+ tool_ends = current_text.count(self.tool_call_end_token)
343
+ if tool_ends > self.current_tool_index:
344
+ # This tool has ended, advance to next
345
+ self.current_tool_index += 1
346
+ self.header_sent = False
347
+ self.param_count = 0
348
+ self.json_started = False
349
+ self.json_closed = False
350
+ self.accumulated_params = {}
351
+
352
+ # Check if there are more tool calls
353
+ tool_starts = current_text.count(self.tool_call_start_token)
354
+ if self.current_tool_index >= tool_starts:
355
+ # No more tool calls
356
+ self.is_tool_call_started = False
357
+ # Continue processing next tool
358
+ return None
359
+
360
+ # Handle normal content before tool calls
361
+ if not self.is_tool_call_started:
362
+ # Check if tool call is starting
363
+ if self.tool_call_start_token_id in delta_token_ids or self.tool_call_start_token in delta_text:
364
+ self.is_tool_call_started = True
365
+ # Return any content before the tool call
366
+ if self.tool_call_start_token in delta_text:
367
+ content_before = delta_text[:delta_text.index(
368
+ self.tool_call_start_token)]
369
+ if content_before:
370
+ return DeltaMessage(content=content_before)
371
+ return None
372
+ else:
373
+ # Check if we're between tool calls - skip whitespace
374
+ if current_text.rstrip().endswith(self.tool_call_end_token):
375
+ # We just ended a tool call, skip whitespace
376
+ if delta_text.strip() == "":
377
+ return None
378
+ # Normal content, no tool call
379
+ return DeltaMessage(content=delta_text)
380
+
381
+ # Check if we're between tool calls (waiting for next one)
382
+ # Count tool calls we've seen vs processed
383
+ tool_starts_count = current_text.count(self.tool_call_start_token)
384
+ if self.current_tool_index >= tool_starts_count:
385
+ # We're past all tool calls, shouldn't be here
386
+ return None
387
+
388
+ # We're in a tool call, find the current tool call portion
389
+ # Need to find the correct tool call based on current_tool_index
390
+ tool_starts = []
391
+ idx = 0
392
+ while True:
393
+ idx = current_text.find(self.tool_call_start_token, idx)
394
+ if idx == -1:
395
+ break
396
+ tool_starts.append(idx)
397
+ idx += len(self.tool_call_start_token)
398
+
399
+ if self.current_tool_index >= len(tool_starts):
400
+ # No more tool calls to process yet
401
+ return None
402
+
403
+ tool_start_idx = tool_starts[self.current_tool_index]
404
+ # Find where this tool call ends (or current position if not ended yet)
405
+ tool_end_idx = current_text.find(self.tool_call_end_token,
406
+ tool_start_idx)
407
+ if tool_end_idx == -1:
408
+ tool_text = current_text[tool_start_idx:]
409
+ else:
410
+ tool_text = current_text[tool_start_idx:tool_end_idx +
411
+ len(self.tool_call_end_token)]
412
+
413
+ # Looking for function header
414
+ if not self.header_sent:
415
+ if self.tool_call_prefix in tool_text:
416
+ func_start = tool_text.find(self.tool_call_prefix) + len(
417
+ self.tool_call_prefix)
418
+ func_end = tool_text.find(">", func_start)
419
+
420
+ if func_end != -1:
421
+ # Found complete function name
422
+ self.current_function_name = tool_text[func_start:func_end]
423
+ self.current_tool_id = self._generate_tool_call_id()
424
+ self.header_sent = True
425
+ self.in_function = True
426
+
427
+ # IMPORTANT: Add to prev_tool_call_arr immediately when we detect a tool call
428
+ # This ensures finish_reason="tool_calls" even if parsing isn't complete
429
+ already_added = any(
430
+ tool.get("name") == self.current_function_name
431
+ for tool in self.prev_tool_call_arr)
432
+ if not already_added:
433
+ self.prev_tool_call_arr.append({
434
+ "name": self.current_function_name,
435
+ "arguments":
436
+ "{}", # Placeholder, will be updated later
437
+ })
438
+
439
+ # Send header with function info
440
+ return DeltaMessage(tool_calls=[
441
+ DeltaToolCall(
442
+ index=self.current_tool_index,
443
+ id=self.current_tool_id,
444
+ function=DeltaFunctionCall(
445
+ name=self.current_function_name, arguments=""),
446
+ type="function",
447
+ )
448
+ ])
449
+ return None
450
+
451
+ # We've sent header, now handle function body
452
+ if self.in_function:
453
+ # Send opening brace if not sent yet
454
+ if not self.json_started and self.parameter_prefix not in delta_text:
455
+ self.json_started = True
456
+ return DeltaMessage(tool_calls=[
457
+ DeltaToolCall(
458
+ index=self.current_tool_index,
459
+ function=DeltaFunctionCall(arguments="{"),
460
+ )
461
+ ])
462
+
463
+ # Make sure json_started is set if we're processing parameters
464
+ if not self.json_started:
465
+ self.json_started = True
466
+
467
+ # Check for function end in accumulated text
468
+ if not self.json_closed and self.function_end_token in tool_text:
469
+ # Close JSON
470
+ self.json_closed = True
471
+
472
+ # Extract the complete tool call to update prev_tool_call_arr with final arguments
473
+ # Find the function content
474
+ func_start = tool_text.find(self.tool_call_prefix) + len(
475
+ self.tool_call_prefix)
476
+ func_content_end = tool_text.find(self.function_end_token,
477
+ func_start)
478
+ if func_content_end != -1:
479
+ func_content = tool_text[func_start:func_content_end]
480
+ # Parse to get the complete arguments
481
+ try:
482
+ parsed_tool = self._parse_xml_function_call(
483
+ func_content, self.streaming_request.tools
484
+ if self.streaming_request else None)
485
+ if parsed_tool:
486
+ # Update existing entry in prev_tool_call_arr with complete arguments
487
+ for i, tool in enumerate(self.prev_tool_call_arr):
488
+ if tool.get(
489
+ "name") == parsed_tool.function.name:
490
+ self.prev_tool_call_arr[i][
491
+ "arguments"] = parsed_tool.function.arguments
492
+ break
493
+ except Exception:
494
+ pass # Ignore parsing errors during streaming
495
+
496
+ result = DeltaMessage(tool_calls=[
497
+ DeltaToolCall(
498
+ index=self.current_tool_index,
499
+ function=DeltaFunctionCall(arguments="}"),
500
+ )
501
+ ])
502
+
503
+ # Reset state for next tool
504
+ self.in_function = False
505
+ self.json_closed = True
506
+ self.accumulated_params = {}
507
+
508
+ return result
509
+
510
+ # Look for parameters
511
+ # Find all parameter starts
512
+ param_starts = []
513
+ idx = 0
514
+ while True:
515
+ idx = tool_text.find(self.parameter_prefix, idx)
516
+ if idx == -1:
517
+ break
518
+ param_starts.append(idx)
519
+ idx += len(self.parameter_prefix)
520
+
521
+ # Check if we should start a new parameter
522
+ if not self.in_param and self.param_count < len(param_starts):
523
+
524
+ if len(param_starts) > self.param_count:
525
+ # Process the next parameter
526
+ param_idx = param_starts[self.param_count]
527
+ param_start = param_idx + len(self.parameter_prefix)
528
+ remaining = tool_text[param_start:]
529
+
530
+ if ">" in remaining:
531
+ # We have the complete parameter name
532
+ name_end = remaining.find(">")
533
+ self.current_param_name = remaining[:name_end]
534
+
535
+ # Find the parameter value
536
+ value_start = param_start + name_end + 1
537
+ value_text = tool_text[value_start:]
538
+ if value_text.startswith("\n"):
539
+ value_text = value_text[1:]
540
+
541
+ # Find where this parameter ends
542
+ param_end_idx = value_text.find(
543
+ self.parameter_end_token)
544
+ if param_end_idx == -1:
545
+ # No closing tag, look for next parameter or function end
546
+ next_param_idx = value_text.find(
547
+ self.parameter_prefix)
548
+ func_end_idx = value_text.find(
549
+ self.function_end_token)
550
+
551
+ if next_param_idx != -1 and (func_end_idx == -1
552
+ or next_param_idx
553
+ < func_end_idx):
554
+ param_end_idx = next_param_idx
555
+ elif func_end_idx != -1:
556
+ param_end_idx = func_end_idx
557
+ else:
558
+ # Neither found, check if tool call is complete
559
+ if self.tool_call_end_token in tool_text:
560
+ # Tool call is complete, so parameter must be complete too
561
+ # Use all remaining text before function end as value
562
+ param_end_idx = len(value_text)
563
+ else:
564
+ # Still streaming, wait for more content
565
+ return None
566
+
567
+ if param_end_idx != -1:
568
+ # Complete parameter found
569
+ param_value = value_text[:param_end_idx]
570
+ if param_value.endswith("\n"):
571
+ param_value = param_value[:-1]
572
+
573
+ # Store raw value for later processing
574
+ self.accumulated_params[
575
+ self.current_param_name] = param_value
576
+
577
+ # Get parameter configuration for type conversion
578
+ param_config = self._get_arguments_config(
579
+ self.current_function_name,
580
+ self.streaming_request.tools
581
+ if self.streaming_request else None)
582
+
583
+ # Convert the parameter value to the appropriate type
584
+ converted_value = self._convert_param_value(
585
+ param_value, self.current_param_name,
586
+ param_config, self.current_function_name)
587
+
588
+ # Build JSON fragment based on the converted type
589
+ # Use json.dumps to properly serialize the value
590
+ serialized_value = json.dumps(converted_value,
591
+ ensure_ascii=False)
592
+
593
+ if self.param_count == 0:
594
+ json_fragment = f'"{self.current_param_name}": {serialized_value}'
595
+ else:
596
+ json_fragment = f', "{self.current_param_name}": {serialized_value}'
597
+
598
+ self.param_count += 1
599
+
600
+ return DeltaMessage(tool_calls=[
601
+ DeltaToolCall(
602
+ index=self.current_tool_index,
603
+ function=DeltaFunctionCall(
604
+ arguments=json_fragment),
605
+ )
606
+ ])
607
+
608
+ # Continue parameter value - Not used in the current implementation
609
+ # since we process complete parameters above
610
+ if self.in_param:
611
+ if self.parameter_end_token in delta_text:
612
+ # End of parameter
613
+ end_idx = delta_text.find(self.parameter_end_token)
614
+ value_chunk = delta_text[:end_idx]
615
+
616
+ # Skip past > if at start
617
+ if not self.current_param_value and ">" in value_chunk:
618
+ gt_idx = value_chunk.find(">")
619
+ value_chunk = value_chunk[gt_idx + 1:]
620
+
621
+ if not self.current_param_value and value_chunk.startswith(
622
+ "\n"):
623
+ value_chunk = value_chunk[1:]
624
+
625
+ # Store complete value
626
+ full_value = self.current_param_value + value_chunk
627
+ self.accumulated_params[
628
+ self.current_param_name] = full_value
629
+
630
+ # Get parameter configuration for type conversion
631
+ param_config = self._get_arguments_config(
632
+ self.current_function_name,
633
+ self.streaming_request.tools
634
+ if self.streaming_request else None)
635
+
636
+ # Convert the parameter value to the appropriate type
637
+ converted_value = self._convert_param_value(
638
+ full_value, self.current_param_name, param_config,
639
+ self.current_function_name)
640
+
641
+ # Serialize the converted value
642
+ serialized_value = json.dumps(converted_value,
643
+ ensure_ascii=False)
644
+
645
+ # Since we've been streaming the quoted version, we need to close it properly
646
+ # This is complex - for now just complete the value
647
+ self.in_param = False
648
+ self.current_param_value = ""
649
+
650
+ # Just close the current parameter string
651
+ return DeltaMessage(tool_calls=[
652
+ DeltaToolCall(
653
+ index=self.current_tool_index,
654
+ function=DeltaFunctionCall(
655
+ arguments='"'), # Close the string quote
656
+ )
657
+ ])
658
+ else:
659
+ # Continue accumulating value
660
+ value_chunk = delta_text
661
+
662
+ # Handle first chunk after param name
663
+ if not self.current_param_value and ">" in value_chunk:
664
+ gt_idx = value_chunk.find(">")
665
+ value_chunk = value_chunk[gt_idx + 1:]
666
+
667
+ if not self.current_param_value and value_chunk.startswith(
668
+ "\n"):
669
+ value_chunk = value_chunk[1:]
670
+
671
+ if value_chunk:
672
+ # Stream the escaped delta
673
+ prev_escaped = json.dumps(
674
+ self.current_param_value, ensure_ascii=False
675
+ )[1:-1] if self.current_param_value else ""
676
+ self.current_param_value += value_chunk
677
+ full_escaped = json.dumps(self.current_param_value,
678
+ ensure_ascii=False)[1:-1]
679
+ delta_escaped = full_escaped[len(prev_escaped):]
680
+
681
+ if delta_escaped:
682
+ return DeltaMessage(tool_calls=[
683
+ DeltaToolCall(
684
+ index=self.current_tool_index,
685
+ function=DeltaFunctionCall(
686
+ arguments=delta_escaped),
687
+ )
688
+ ])
689
+
690
+ return None
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
3
+ size 11422650
tokenizer_config.json ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "backend": "tokenizers",
4
+ "bos_token": null,
5
+ "clean_up_tokenization_spaces": false,
6
+ "eos_token": "<|im_end|>",
7
+ "errors": "replace",
8
+ "extra_special_tokens": [
9
+ "<|im_start|>",
10
+ "<|im_end|>",
11
+ "<|object_ref_start|>",
12
+ "<|object_ref_end|>",
13
+ "<|box_start|>",
14
+ "<|box_end|>",
15
+ "<|quad_start|>",
16
+ "<|quad_end|>",
17
+ "<|vision_start|>",
18
+ "<|vision_end|>",
19
+ "<|vision_pad|>",
20
+ "<|image_pad|>",
21
+ "<|video_pad|>"
22
+ ],
23
+ "is_local": true,
24
+ "model_max_length": 1048576,
25
+ "pad_token": "<|endoftext|>",
26
+ "split_special_tokens": false,
27
+ "tokenizer_class": "Qwen2Tokenizer",
28
+ "tool_parser_type": "qwen3_coder",
29
+ "unk_token": null
30
+ }