Hugging Face's logo

INC4AI
/
LongCat-Flash-Thinking-2601-int4-mixed-AutoRound

Text Generation
Safetensors
longcat
conversational
custom_code
4-bit precision
auto-round
Model card Files
xet
 Community
INC4AI commited on
Commit
6148d4c
·
verified ·
1 Parent(s): 9b3d44b

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. chat_template.jinja +138 -0
  3. config.json +3401 -0
  4. configuration_longcat.py +210 -0
  5. generation_config.json +7 -0
  6. model-00001-of-00061.safetensors +3 -0
  7. model-00002-of-00061.safetensors +3 -0
  8. model-00003-of-00061.safetensors +3 -0
  9. model-00004-of-00061.safetensors +3 -0
  10. model-00005-of-00061.safetensors +3 -0
  11. model-00006-of-00061.safetensors +3 -0
  12. model-00007-of-00061.safetensors +3 -0
  13. model-00008-of-00061.safetensors +3 -0
  14. model-00009-of-00061.safetensors +3 -0
  15. model-00010-of-00061.safetensors +3 -0
  16. model-00011-of-00061.safetensors +3 -0
  17. model-00012-of-00061.safetensors +3 -0
  18. model-00013-of-00061.safetensors +3 -0
  19. model-00014-of-00061.safetensors +3 -0
  20. model-00015-of-00061.safetensors +3 -0
  21. model-00016-of-00061.safetensors +3 -0
  22. model-00017-of-00061.safetensors +3 -0
  23. model-00018-of-00061.safetensors +3 -0
  24. model-00019-of-00061.safetensors +3 -0
  25. model-00020-of-00061.safetensors +3 -0
  26. model-00021-of-00061.safetensors +3 -0
  27. model-00022-of-00061.safetensors +3 -0
  28. model-00023-of-00061.safetensors +3 -0
  29. model-00024-of-00061.safetensors +3 -0
  30. model-00025-of-00061.safetensors +3 -0
  31. model-00026-of-00061.safetensors +3 -0
  32. model-00027-of-00061.safetensors +3 -0
  33. model-00028-of-00061.safetensors +3 -0
  34. model-00029-of-00061.safetensors +3 -0
  35. model-00030-of-00061.safetensors +3 -0
  36. model-00031-of-00061.safetensors +3 -0
  37. model-00032-of-00061.safetensors +3 -0
  38. model-00033-of-00061.safetensors +3 -0
  39. model-00034-of-00061.safetensors +3 -0
  40. model-00035-of-00061.safetensors +3 -0
  41. model-00036-of-00061.safetensors +3 -0
  42. model-00037-of-00061.safetensors +3 -0
  43. model-00038-of-00061.safetensors +3 -0
  44. model-00039-of-00061.safetensors +3 -0
  45. model-00040-of-00061.safetensors +3 -0
  46. model-00041-of-00061.safetensors +3 -0
  47. model-00042-of-00061.safetensors +3 -0
  48. model-00043-of-00061.safetensors +3 -0
  49. model-00044-of-00061.safetensors +3 -0
  50. model-00045-of-00061.safetensors +3 -0
.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
+ model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
chat_template.jinja ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- set tool_choice = tool_choice | default('auto') %}
2
+ {%- set enable_thinking = enable_thinking | default(none) %}
3
+ {%- set save_history_reasoning_content = save_history_reasoning_content | default(false) %}
4
+ {%- set ns = namespace(tool_types = [], last_query_index = -1) %}
5
+ {%- for idx in range(messages|length - 1) %}
6
+ {%- set msg = messages[idx] %}
7
+ {%- if msg.role == 'assistant' and not msg.tool_calls %}
8
+ {%- set ns.last_query_index = idx %}
9
+ {%- endif %}
10
+ {%- endfor %}
11
+
12
+ {%- if tools and tool_choice != 'none' %}
13
+ {{- "<longcat_tool_declare>\n"-}}
14
+ {{- "# Tools\n" }}
15
+ {{- "You have access to the following tools:\n\n" }}
16
+ {%- for tool in tools %}
17
+ {%- if tool.type not in ns.tool_types %}
18
+ {%- set ns.tool_types = ns.tool_types + [tool.type] %}
19
+ {{- "## Tool namespace: " ~ tool.type ~ "\n\n" }}
20
+ {%- endif %}
21
+ {%- if tool.type == 'code_interpreter' %}
22
+ {%- set tool = {"type":"code_interpreter","function":{"name":"code_interpreter_preview","description":"The code will be executed in a stateful Jupyter notebook sandbox environment, only supports local computation, data processing, and file operations.\nCode sandbox environment (network isolated) Any external network requests or online API calls are prohibited.\nIf online functionality is needed, please use other permitted tools.\nCode will respond with the output of the execution or time out after 60.0 seconds. ","parameters":{"type":"object","properties":{"language":{"type":"string","description":"The programming language of the code to be executed. Available values: python (Default), java, go, js, ts, c, c++."},"code":{"type":"string","description":"Python code to be executed must not include the following:\n- Importing network libraries such as requests, httplib, etc.\n- Any form of HTTP requests.\n- External API calls.\n- Network port operations. Example: ```python\nimport pandas as pd\npd.DataFrame({'A':[1,2]})\n```"},"timeout":{"type":"number","description":"The maximum execution time of the code, in seconds. Default is 60.0."}}},"required":["code"]}} %}
23
+ {%- endif %}
24
+ {{- "### Tool name: " + tool.function.name + "\n" }}
25
+ {{- "Description: " + tool.function.description + "\n\n" }}
26
+ {{- "InputSchema: " + tool.function.parameters | tojson(ensure_ascii=False) + "\n\n" }}
27
+ {%- endfor %}
28
+ {{- '**Note**: For each function call, output the function name and arguments within the following XML format:\n<longcat_tool_call>{function-name}\n<longcat_arg_key>{arg-key-1}</longcat_arg_key>\n<longcat_arg_value>{arg-value-1}</longcat_arg_value>\n<longcat_arg_key>{arg-key-2}</longcat_arg_key>\n<longcat_arg_value>{arg-value-2}</longcat_arg_value>\n...\n</longcat_tool_call>\n' }}
29
+ {{- "</longcat_tool_declare>"-}}
30
+ {%- endif %}
31
+
32
+ {%- for msg in messages %}
33
+ {%- if msg.role == "system" %}
34
+ {{- "<longcat_system>" + msg.content }}
35
+ {%- elif msg.role == "user" %}
36
+ {{- "<longcat_user>" }}
37
+ {%- if msg["files"] %}
38
+ {{- '<longcat_files>\n' ~ msg.files | tojson(indent=2) ~ '\n</longcat_files>' }}
39
+ {%- endif %}
40
+ {{- msg.content }}
41
+ {%- if save_history_reasoning_content and enable_thinking is not none %}
42
+ {%- if loop.index0 == messages|length - 1 %}
43
+ {# pass #}
44
+ {%- else %}
45
+ {%- if enable_thinking == true %}
46
+ {{- " /think_on" }}
47
+ {%- if thinking_budget %}
48
+ {%- if thinking_budget < 1024 %}
49
+ {%- set thinking_budget = 1024 %}
50
+ {%- endif %}
51
+ {{- "\nthinking_budget: < " ~ thinking_budget ~ "." }}
52
+ {%- endif %}
53
+ {{- " " }}
54
+ {%- elif enable_thinking == false %}
55
+ {{- " /think_off " }}
56
+ {%- endif %}
57
+ {%- endif %}
58
+ {%- elif loop.index0 > ns.last_query_index and enable_thinking is not none %}
59
+ {%- if loop.index0 == messages|length - 1 %}
60
+ {# pass #}
61
+ {%- else %}
62
+ {%- if enable_thinking == true %}
63
+ {{- " /think_on" }}
64
+ {%- if thinking_budget %}
65
+ {%- if thinking_budget < 1024 %}
66
+ {%- set thinking_budget = 1024 %}
67
+ {%- endif %}
68
+ {{- "\nthinking_budget: < " ~ thinking_budget ~ "." }}
69
+ {%- endif %}
70
+ {{- " " }}
71
+ {%- elif enable_thinking == false %}
72
+ {{- " /think_off " }}
73
+ {%- endif %}
74
+ {%- endif %}
75
+ {%- endif %}
76
+ {%- elif msg.role == "assistant" %}
77
+ {{- "<longcat_assistant>" }}
78
+ {%- if save_history_reasoning_content == true or loop.index0 > ns.last_query_index %}
79
+ {%- if enable_thinking == true and msg.reasoning_content%}
80
+ {{- "<longcat_think>\n" ~ msg.reasoning_content ~ "\n</longcat_think>\n" }}
81
+ {%- elif enable_thinking == false %}
82
+ {{- "<longcat_think>\n\n</longcat_think>\n" }}
83
+ {%- endif %}
84
+ {%- endif %}
85
+ {%- if msg.content %}
86
+ {{- msg.content }}
87
+ {%- endif %}
88
+ {%- if msg.tool_calls %}
89
+ {%- for tool_call in msg.tool_calls -%}
90
+ {{- "<longcat_tool_call>" ~ tool_call.function.name ~ "\n" -}}
91
+ {% set _args = tool_call.function.arguments %}
92
+ {% for k, v in _args.items() %}
93
+ {{- "<longcat_arg_key>" ~ k ~ "</longcat_arg_key>\n" -}}
94
+ {{- "<longcat_arg_value>" ~ (v if v is string else v | tojson(ensure_ascii=False)) ~ "</longcat_arg_value>\n" -}}
95
+ {% endfor %}
96
+ {{- "</longcat_tool_call>\n" }}
97
+ {%- endfor %}
98
+ {%- endif %}
99
+ {{- "</longcat_s>" -}}
100
+ {%- elif msg.role == "tool" %}
101
+ {%- if messages[loop.index0 - 1].role != "tool" %}
102
+ {{- "<longcat_observation>" -}}
103
+ {%- endif %}
104
+ {{- "<longcat_tool_response>" ~ msg.name ~ "\n" ~ msg.content ~ "</longcat_tool_response>"-}}
105
+ {%- if loop.index0 == messages|length - 1 %}
106
+ {# pass #}
107
+ {%- elif save_history_reasoning_content or loop.index0 > ns.last_query_index %}
108
+ {%- if enable_thinking == true and messages[loop.index0 + 1].role != "tool" %}
109
+ {{- " /think_on" }}
110
+ {%- if thinking_budget %}
111
+ {%- if thinking_budget < 1024 %}
112
+ {%- set thinking_budget = 1024 %}
113
+ {%- endif %}
114
+ {{- "\nthinking_budget: < " ~ thinking_budget ~ "." }}
115
+ {%- endif %}
116
+ {{- " " }}
117
+ {%- elif enable_thinking == false %}
118
+ {{- " /think_off " }}
119
+ {%- endif %}
120
+ {%- endif %}
121
+ {%- endif %}
122
+ {%- endfor %}
123
+ {%- if add_generation_prompt %}
124
+ {%- if enable_thinking == true %}
125
+ {{- " /think_on" }}
126
+ {%- if thinking_budget %}
127
+ {%- if thinking_budget < 1024 %}
128
+ {%- set thinking_budget = 1024 %}
129
+ {%- endif %}
130
+ {{- "\nthinking_budget: < " ~ thinking_budget ~ "." }}
131
+ {%- endif %}
132
+ {{- " <longcat_assistant><longcat_think>\n" }}
133
+ {%- elif enable_thinking == false %}
134
+ {{- " /think_off <longcat_assistant><longcat_think>\n\n</longcat_think>\n" }}
135
+ {%- else %}
136
+ {{- "<longcat_assistant>" }}
137
+ {%- endif %}
138
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,3401 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "LongcatForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "attention_method": "MLA",
8
+ "auto_map": {
9
+ "AutoConfig": "configuration_longcat.LongcatConfig",
10
+ "AutoModel": "modeling_longcat.LongcatModel",
11
+ "AutoModelForCausalLM": "modeling_longcat.LongcatForCausalLM"
12
+ },
13
+ "bos_token_id": 1,
14
+ "eos_token_id": 2,
15
+ "expert_ffn_hidden_size": 2048,
16
+ "ffn_hidden_size": 12288,
17
+ "head_dim": 64,
18
+ "hidden_act": "silu",
19
+ "hidden_size": 6144,
20
+ "initializer_range": 0.006,
21
+ "kv_lora_rank": 512,
22
+ "max_position_embeddings": 983040,
23
+ "mla_scale_kv_lora": true,
24
+ "mla_scale_q_lora": true,
25
+ "model_type": "longcat",
26
+ "moe_topk": 12,
27
+ "n_routed_experts": 512,
28
+ "norm_topk_prob": false,
29
+ "num_attention_heads": 64,
30
+ "num_key_value_heads": 64,
31
+ "num_layers": 28,
32
+ "q_lora_rank": 1536,
33
+ "qk_head_dim": 192,
34
+ "qk_nope_head_dim": 128,
35
+ "qk_rope_head_dim": 64,
36
+ "quantization_config": {
37
+ "autoround_version": "0.9.5",
38
+ "bits": 4,
39
+ "data_type": "int",
40
+ "extra_config": {
41
+ "model.layers.0.mlp.router.classifier": {
42
+ "act_bits": 16,
43
+ "bits": 8,
44
+ "data_type": "int",
45
+ "group_size": 128,
46
+ "sym": true
47
+ },
48
+ "model.layers.0.mlps.0.down_proj": {
49
+ "act_bits": 16,
50
+ "bits": 8,
51
+ "data_type": "int",
52
+ "group_size": 128,
53
+ "sym": true
54
+ },
55
+ "model.layers.0.mlps.0.gate_proj": {
56
+ "act_bits": 16,
57
+ "bits": 8,
58
+ "data_type": "int",
59
+ "group_size": 128,
60
+ "sym": true
61
+ },
62
+ "model.layers.0.mlps.0.up_proj": {
63
+ "act_bits": 16,
64
+ "bits": 8,
65
+ "data_type": "int",
66
+ "group_size": 128,
67
+ "sym": true
68
+ },
69
+ "model.layers.0.mlps.1.down_proj": {
70
+ "act_bits": 16,
71
+ "bits": 8,
72
+ "data_type": "int",
73
+ "group_size": 128,
74
+ "sym": true
75
+ },
76
+ "model.layers.0.mlps.1.gate_proj": {
77
+ "act_bits": 16,
78
+ "bits": 8,
79
+ "data_type": "int",
80
+ "group_size": 128,
81
+ "sym": true
82
+ },
83
+ "model.layers.0.mlps.1.up_proj": {
84
+ "act_bits": 16,
85
+ "bits": 8,
86
+ "data_type": "int",
87
+ "group_size": 128,
88
+ "sym": true
89
+ },
90
+ "model.layers.0.self_attn.0.kv_a_proj_with_mqa": {
91
+ "act_bits": 16,
92
+ "bits": 8,
93
+ "data_type": "int",
94
+ "group_size": 128,
95
+ "sym": true
96
+ },
97
+ "model.layers.0.self_attn.0.kv_b_proj": {
98
+ "act_bits": 16,
99
+ "bits": 8,
100
+ "data_type": "int",
101
+ "group_size": 128,
102
+ "sym": true
103
+ },
104
+ "model.layers.0.self_attn.0.o_proj": {
105
+ "act_bits": 16,
106
+ "bits": 8,
107
+ "data_type": "int",
108
+ "group_size": 128,
109
+ "sym": true
110
+ },
111
+ "model.layers.0.self_attn.0.q_a_proj": {
112
+ "act_bits": 16,
113
+ "bits": 8,
114
+ "data_type": "int",
115
+ "group_size": 128,
116
+ "sym": true
117
+ },
118
+ "model.layers.0.self_attn.0.q_b_proj": {
119
+ "act_bits": 16,
120
+ "bits": 8,
121
+ "data_type": "int",
122
+ "group_size": 128,
123
+ "sym": true
124
+ },
125
+ "model.layers.0.self_attn.1.kv_a_proj_with_mqa": {
126
+ "act_bits": 16,
127
+ "bits": 8,
128
+ "data_type": "int",
129
+ "group_size": 128,
130
+ "sym": true
131
+ },
132
+ "model.layers.0.self_attn.1.kv_b_proj": {
133
+ "act_bits": 16,
134
+ "bits": 8,
135
+ "data_type": "int",
136
+ "group_size": 128,
137
+ "sym": true
138
+ },
139
+ "model.layers.0.self_attn.1.o_proj": {
140
+ "act_bits": 16,
141
+ "bits": 8,
142
+ "data_type": "int",
143
+ "group_size": 128,
144
+ "sym": true
145
+ },
146
+ "model.layers.0.self_attn.1.q_a_proj": {
147
+ "act_bits": 16,
148
+ "bits": 8,
149
+ "data_type": "int",
150
+ "group_size": 128,
151
+ "sym": true
152
+ },
153
+ "model.layers.0.self_attn.1.q_b_proj": {
154
+ "act_bits": 16,
155
+ "bits": 8,
156
+ "data_type": "int",
157
+ "group_size": 128,
158
+ "sym": true
159
+ },
160
+ "model.layers.1.mlp.router.classifier": {
161
+ "act_bits": 16,
162
+ "bits": 8,
163
+ "data_type": "int",
164
+ "group_size": 128,
165
+ "sym": true
166
+ },
167
+ "model.layers.1.mlps.0.down_proj": {
168
+ "act_bits": 16,
169
+ "bits": 8,
170
+ "data_type": "int",
171
+ "group_size": 128,
172
+ "sym": true
173
+ },
174
+ "model.layers.1.mlps.0.gate_proj": {
175
+ "act_bits": 16,
176
+ "bits": 8,
177
+ "data_type": "int",
178
+ "group_size": 128,
179
+ "sym": true
180
+ },
181
+ "model.layers.1.mlps.0.up_proj": {
182
+ "act_bits": 16,
183
+ "bits": 8,
184
+ "data_type": "int",
185
+ "group_size": 128,
186
+ "sym": true
187
+ },
188
+ "model.layers.1.mlps.1.down_proj": {
189
+ "act_bits": 16,
190
+ "bits": 8,
191
+ "data_type": "int",
192
+ "group_size": 128,
193
+ "sym": true
194
+ },
195
+ "model.layers.1.mlps.1.gate_proj": {
196
+ "act_bits": 16,
197
+ "bits": 8,
198
+ "data_type": "int",
199
+ "group_size": 128,
200
+ "sym": true
201
+ },
202
+ "model.layers.1.mlps.1.up_proj": {
203
+ "act_bits": 16,
204
+ "bits": 8,
205
+ "data_type": "int",
206
+ "group_size": 128,
207
+ "sym": true
208
+ },
209
+ "model.layers.1.self_attn.0.kv_a_proj_with_mqa": {
210
+ "act_bits": 16,
211
+ "bits": 8,
212
+ "data_type": "int",
213
+ "group_size": 128,
214
+ "sym": true
215
+ },
216
+ "model.layers.1.self_attn.0.kv_b_proj": {
217
+ "act_bits": 16,
218
+ "bits": 8,
219
+ "data_type": "int",
220
+ "group_size": 128,
221
+ "sym": true
222
+ },
223
+ "model.layers.1.self_attn.0.o_proj": {
224
+ "act_bits": 16,
225
+ "bits": 8,
226
+ "data_type": "int",
227
+ "group_size": 128,
228
+ "sym": true
229
+ },
230
+ "model.layers.1.self_attn.0.q_a_proj": {
231
+ "act_bits": 16,
232
+ "bits": 8,
233
+ "data_type": "int",
234
+ "group_size": 128,
235
+ "sym": true
236
+ },
237
+ "model.layers.1.self_attn.0.q_b_proj": {
238
+ "act_bits": 16,
239
+ "bits": 8,
240
+ "data_type": "int",
241
+ "group_size": 128,
242
+ "sym": true
243
+ },
244
+ "model.layers.1.self_attn.1.kv_a_proj_with_mqa": {
245
+ "act_bits": 16,
246
+ "bits": 8,
247
+ "data_type": "int",
248
+ "group_size": 128,
249
+ "sym": true
250
+ },
251
+ "model.layers.1.self_attn.1.kv_b_proj": {
252
+ "act_bits": 16,
253
+ "bits": 8,
254
+ "data_type": "int",
255
+ "group_size": 128,
256
+ "sym": true
257
+ },
258
+ "model.layers.1.self_attn.1.o_proj": {
259
+ "act_bits": 16,
260
+ "bits": 8,
261
+ "data_type": "int",
262
+ "group_size": 128,
263
+ "sym": true
264
+ },
265
+ "model.layers.1.self_attn.1.q_a_proj": {
266
+ "act_bits": 16,
267
+ "bits": 8,
268
+ "data_type": "int",
269
+ "group_size": 128,
270
+ "sym": true
271
+ },
272
+ "model.layers.1.self_attn.1.q_b_proj": {
273
+ "act_bits": 16,
274
+ "bits": 8,
275
+ "data_type": "int",
276
+ "group_size": 128,
277
+ "sym": true
278
+ },
279
+ "model.layers.10.mlp.router.classifier": {
280
+ "act_bits": 16,
281
+ "bits": 8,
282
+ "data_type": "int",
283
+ "group_size": 128,
284
+ "sym": true
285
+ },
286
+ "model.layers.10.mlps.0.down_proj": {
287
+ "act_bits": 16,
288
+ "bits": 8,
289
+ "data_type": "int",
290
+ "group_size": 128,
291
+ "sym": true
292
+ },
293
+ "model.layers.10.mlps.0.gate_proj": {
294
+ "act_bits": 16,
295
+ "bits": 8,
296
+ "data_type": "int",
297
+ "group_size": 128,
298
+ "sym": true
299
+ },
300
+ "model.layers.10.mlps.0.up_proj": {
301
+ "act_bits": 16,
302
+ "bits": 8,
303
+ "data_type": "int",
304
+ "group_size": 128,
305
+ "sym": true
306
+ },
307
+ "model.layers.10.mlps.1.down_proj": {
308
+ "act_bits": 16,
309
+ "bits": 8,
310
+ "data_type": "int",
311
+ "group_size": 128,
312
+ "sym": true
313
+ },
314
+ "model.layers.10.mlps.1.gate_proj": {
315
+ "act_bits": 16,
316
+ "bits": 8,
317
+ "data_type": "int",
318
+ "group_size": 128,
319
+ "sym": true
320
+ },
321
+ "model.layers.10.mlps.1.up_proj": {
322
+ "act_bits": 16,
323
+ "bits": 8,
324
+ "data_type": "int",
325
+ "group_size": 128,
326
+ "sym": true
327
+ },
328
+ "model.layers.10.self_attn.0.kv_a_proj_with_mqa": {
329
+ "act_bits": 16,
330
+ "bits": 8,
331
+ "data_type": "int",
332
+ "group_size": 128,
333
+ "sym": true
334
+ },
335
+ "model.layers.10.self_attn.0.kv_b_proj": {
336
+ "act_bits": 16,
337
+ "bits": 8,
338
+ "data_type": "int",
339
+ "group_size": 128,
340
+ "sym": true
341
+ },
342
+ "model.layers.10.self_attn.0.o_proj": {
343
+ "act_bits": 16,
344
+ "bits": 8,
345
+ "data_type": "int",
346
+ "group_size": 128,
347
+ "sym": true
348
+ },
349
+ "model.layers.10.self_attn.0.q_a_proj": {
350
+ "act_bits": 16,
351
+ "bits": 8,
352
+ "data_type": "int",
353
+ "group_size": 128,
354
+ "sym": true
355
+ },
356
+ "model.layers.10.self_attn.0.q_b_proj": {
357
+ "act_bits": 16,
358
+ "bits": 8,
359
+ "data_type": "int",
360
+ "group_size": 128,
361
+ "sym": true
362
+ },
363
+ "model.layers.10.self_attn.1.kv_a_proj_with_mqa": {
364
+ "act_bits": 16,
365
+ "bits": 8,
366
+ "data_type": "int",
367
+ "group_size": 128,
368
+ "sym": true
369
+ },
370
+ "model.layers.10.self_attn.1.kv_b_proj": {
371
+ "act_bits": 16,
372
+ "bits": 8,
373
+ "data_type": "int",
374
+ "group_size": 128,
375
+ "sym": true
376
+ },
377
+ "model.layers.10.self_attn.1.o_proj": {
378
+ "act_bits": 16,
379
+ "bits": 8,
380
+ "data_type": "int",
381
+ "group_size": 128,
382
+ "sym": true
383
+ },
384
+ "model.layers.10.self_attn.1.q_a_proj": {
385
+ "act_bits": 16,
386
+ "bits": 8,
387
+ "data_type": "int",
388
+ "group_size": 128,
389
+ "sym": true
390
+ },
391
+ "model.layers.10.self_attn.1.q_b_proj": {
392
+ "act_bits": 16,
393
+ "bits": 8,
394
+ "data_type": "int",
395
+ "group_size": 128,
396
+ "sym": true
397
+ },
398
+ "model.layers.11.mlp.router.classifier": {
399
+ "act_bits": 16,
400
+ "bits": 8,
401
+ "data_type": "int",
402
+ "group_size": 128,
403
+ "sym": true
404
+ },
405
+ "model.layers.11.mlps.0.down_proj": {
406
+ "act_bits": 16,
407
+ "bits": 8,
408
+ "data_type": "int",
409
+ "group_size": 128,
410
+ "sym": true
411
+ },
412
+ "model.layers.11.mlps.0.gate_proj": {
413
+ "act_bits": 16,
414
+ "bits": 8,
415
+ "data_type": "int",
416
+ "group_size": 128,
417
+ "sym": true
418
+ },
419
+ "model.layers.11.mlps.0.up_proj": {
420
+ "act_bits": 16,
421
+ "bits": 8,
422
+ "data_type": "int",
423
+ "group_size": 128,
424
+ "sym": true
425
+ },
426
+ "model.layers.11.mlps.1.down_proj": {
427
+ "act_bits": 16,
428
+ "bits": 8,
429
+ "data_type": "int",
430
+ "group_size": 128,
431
+ "sym": true
432
+ },
433
+ "model.layers.11.mlps.1.gate_proj": {
434
+ "act_bits": 16,
435
+ "bits": 8,
436
+ "data_type": "int",
437
+ "group_size": 128,
438
+ "sym": true
439
+ },
440
+ "model.layers.11.mlps.1.up_proj": {
441
+ "act_bits": 16,
442
+ "bits": 8,
443
+ "data_type": "int",
444
+ "group_size": 128,
445
+ "sym": true
446
+ },
447
+ "model.layers.11.self_attn.0.kv_a_proj_with_mqa": {
448
+ "act_bits": 16,
449
+ "bits": 8,
450
+ "data_type": "int",
451
+ "group_size": 128,
452
+ "sym": true
453
+ },
454
+ "model.layers.11.self_attn.0.kv_b_proj": {
455
+ "act_bits": 16,
456
+ "bits": 8,
457
+ "data_type": "int",
458
+ "group_size": 128,
459
+ "sym": true
460
+ },
461
+ "model.layers.11.self_attn.0.o_proj": {
462
+ "act_bits": 16,
463
+ "bits": 8,
464
+ "data_type": "int",
465
+ "group_size": 128,
466
+ "sym": true
467
+ },
468
+ "model.layers.11.self_attn.0.q_a_proj": {
469
+ "act_bits": 16,
470
+ "bits": 8,
471
+ "data_type": "int",
472
+ "group_size": 128,
473
+ "sym": true
474
+ },
475
+ "model.layers.11.self_attn.0.q_b_proj": {
476
+ "act_bits": 16,
477
+ "bits": 8,
478
+ "data_type": "int",
479
+ "group_size": 128,
480
+ "sym": true
481
+ },
482
+ "model.layers.11.self_attn.1.kv_a_proj_with_mqa": {
483
+ "act_bits": 16,
484
+ "bits": 8,
485
+ "data_type": "int",
486
+ "group_size": 128,
487
+ "sym": true
488
+ },
489
+ "model.layers.11.self_attn.1.kv_b_proj": {
490
+ "act_bits": 16,
491
+ "bits": 8,
492
+ "data_type": "int",
493
+ "group_size": 128,
494
+ "sym": true
495
+ },
496
+ "model.layers.11.self_attn.1.o_proj": {
497
+ "act_bits": 16,
498
+ "bits": 8,
499
+ "data_type": "int",
500
+ "group_size": 128,
501
+ "sym": true
502
+ },
503
+ "model.layers.11.self_attn.1.q_a_proj": {
504
+ "act_bits": 16,
505
+ "bits": 8,
506
+ "data_type": "int",
507
+ "group_size": 128,
508
+ "sym": true
509
+ },
510
+ "model.layers.11.self_attn.1.q_b_proj": {
511
+ "act_bits": 16,
512
+ "bits": 8,
513
+ "data_type": "int",
514
+ "group_size": 128,
515
+ "sym": true
516
+ },
517
+ "model.layers.12.mlp.router.classifier": {
518
+ "act_bits": 16,
519
+ "bits": 8,
520
+ "data_type": "int",
521
+ "group_size": 128,
522
+ "sym": true
523
+ },
524
+ "model.layers.12.mlps.0.down_proj": {
525
+ "act_bits": 16,
526
+ "bits": 8,
527
+ "data_type": "int",
528
+ "group_size": 128,
529
+ "sym": true
530
+ },
531
+ "model.layers.12.mlps.0.gate_proj": {
532
+ "act_bits": 16,
533
+ "bits": 8,
534
+ "data_type": "int",
535
+ "group_size": 128,
536
+ "sym": true
537
+ },
538
+ "model.layers.12.mlps.0.up_proj": {
539
+ "act_bits": 16,
540
+ "bits": 8,
541
+ "data_type": "int",
542
+ "group_size": 128,
543
+ "sym": true
544
+ },
545
+ "model.layers.12.mlps.1.down_proj": {
546
+ "act_bits": 16,
547
+ "bits": 8,
548
+ "data_type": "int",
549
+ "group_size": 128,
550
+ "sym": true
551
+ },
552
+ "model.layers.12.mlps.1.gate_proj": {
553
+ "act_bits": 16,
554
+ "bits": 8,
555
+ "data_type": "int",
556
+ "group_size": 128,
557
+ "sym": true
558
+ },
559
+ "model.layers.12.mlps.1.up_proj": {
560
+ "act_bits": 16,
561
+ "bits": 8,
562
+ "data_type": "int",
563
+ "group_size": 128,
564
+ "sym": true
565
+ },
566
+ "model.layers.12.self_attn.0.kv_a_proj_with_mqa": {
567
+ "act_bits": 16,
568
+ "bits": 8,
569
+ "data_type": "int",
570
+ "group_size": 128,
571
+ "sym": true
572
+ },
573
+ "model.layers.12.self_attn.0.kv_b_proj": {
574
+ "act_bits": 16,
575
+ "bits": 8,
576
+ "data_type": "int",
577
+ "group_size": 128,
578
+ "sym": true
579
+ },
580
+ "model.layers.12.self_attn.0.o_proj": {
581
+ "act_bits": 16,
582
+ "bits": 8,
583
+ "data_type": "int",
584
+ "group_size": 128,
585
+ "sym": true
586
+ },
587
+ "model.layers.12.self_attn.0.q_a_proj": {
588
+ "act_bits": 16,
589
+ "bits": 8,
590
+ "data_type": "int",
591
+ "group_size": 128,
592
+ "sym": true
593
+ },
594
+ "model.layers.12.self_attn.0.q_b_proj": {
595
+ "act_bits": 16,
596
+ "bits": 8,
597
+ "data_type": "int",
598
+ "group_size": 128,
599
+ "sym": true
600
+ },
601
+ "model.layers.12.self_attn.1.kv_a_proj_with_mqa": {
602
+ "act_bits": 16,
603
+ "bits": 8,
604
+ "data_type": "int",
605
+ "group_size": 128,
606
+ "sym": true
607
+ },
608
+ "model.layers.12.self_attn.1.kv_b_proj": {
609
+ "act_bits": 16,
610
+ "bits": 8,
611
+ "data_type": "int",
612
+ "group_size": 128,
613
+ "sym": true
614
+ },
615
+ "model.layers.12.self_attn.1.o_proj": {
616
+ "act_bits": 16,
617
+ "bits": 8,
618
+ "data_type": "int",
619
+ "group_size": 128,
620
+ "sym": true
621
+ },
622
+ "model.layers.12.self_attn.1.q_a_proj": {
623
+ "act_bits": 16,
624
+ "bits": 8,
625
+ "data_type": "int",
626
+ "group_size": 128,
627
+ "sym": true
628
+ },
629
+ "model.layers.12.self_attn.1.q_b_proj": {
630
+ "act_bits": 16,
631
+ "bits": 8,
632
+ "data_type": "int",
633
+ "group_size": 128,
634
+ "sym": true
635
+ },
636
+ "model.layers.13.mlp.router.classifier": {
637
+ "act_bits": 16,
638
+ "bits": 8,
639
+ "data_type": "int",
640
+ "group_size": 128,
641
+ "sym": true
642
+ },
643
+ "model.layers.13.mlps.0.down_proj": {
644
+ "act_bits": 16,
645
+ "bits": 8,
646
+ "data_type": "int",
647
+ "group_size": 128,
648
+ "sym": true
649
+ },
650
+ "model.layers.13.mlps.0.gate_proj": {
651
+ "act_bits": 16,
652
+ "bits": 8,
653
+ "data_type": "int",
654
+ "group_size": 128,
655
+ "sym": true
656
+ },
657
+ "model.layers.13.mlps.0.up_proj": {
658
+ "act_bits": 16,
659
+ "bits": 8,
660
+ "data_type": "int",
661
+ "group_size": 128,
662
+ "sym": true
663
+ },
664
+ "model.layers.13.mlps.1.down_proj": {
665
+ "act_bits": 16,
666
+ "bits": 8,
667
+ "data_type": "int",
668
+ "group_size": 128,
669
+ "sym": true
670
+ },
671
+ "model.layers.13.mlps.1.gate_proj": {
672
+ "act_bits": 16,
673
+ "bits": 8,
674
+ "data_type": "int",
675
+ "group_size": 128,
676
+ "sym": true
677
+ },
678
+ "model.layers.13.mlps.1.up_proj": {
679
+ "act_bits": 16,
680
+ "bits": 8,
681
+ "data_type": "int",
682
+ "group_size": 128,
683
+ "sym": true
684
+ },
685
+ "model.layers.13.self_attn.0.kv_a_proj_with_mqa": {
686
+ "act_bits": 16,
687
+ "bits": 8,
688
+ "data_type": "int",
689
+ "group_size": 128,
690
+ "sym": true
691
+ },
692
+ "model.layers.13.self_attn.0.kv_b_proj": {
693
+ "act_bits": 16,
694
+ "bits": 8,
695
+ "data_type": "int",
696
+ "group_size": 128,
697
+ "sym": true
698
+ },
699
+ "model.layers.13.self_attn.0.o_proj": {
700
+ "act_bits": 16,
701
+ "bits": 8,
702
+ "data_type": "int",
703
+ "group_size": 128,
704
+ "sym": true
705
+ },
706
+ "model.layers.13.self_attn.0.q_a_proj": {
707
+ "act_bits": 16,
708
+ "bits": 8,
709
+ "data_type": "int",
710
+ "group_size": 128,
711
+ "sym": true
712
+ },
713
+ "model.layers.13.self_attn.0.q_b_proj": {
714
+ "act_bits": 16,
715
+ "bits": 8,
716
+ "data_type": "int",
717
+ "group_size": 128,
718
+ "sym": true
719
+ },
720
+ "model.layers.13.self_attn.1.kv_a_proj_with_mqa": {
721
+ "act_bits": 16,
722
+ "bits": 8,
723
+ "data_type": "int",
724
+ "group_size": 128,
725
+ "sym": true
726
+ },
727
+ "model.layers.13.self_attn.1.kv_b_proj": {
728
+ "act_bits": 16,
729
+ "bits": 8,
730
+ "data_type": "int",
731
+ "group_size": 128,
732
+ "sym": true
733
+ },
734
+ "model.layers.13.self_attn.1.o_proj": {
735
+ "act_bits": 16,
736
+ "bits": 8,
737
+ "data_type": "int",
738
+ "group_size": 128,
739
+ "sym": true
740
+ },
741
+ "model.layers.13.self_attn.1.q_a_proj": {
742
+ "act_bits": 16,
743
+ "bits": 8,
744
+ "data_type": "int",
745
+ "group_size": 128,
746
+ "sym": true
747
+ },
748
+ "model.layers.13.self_attn.1.q_b_proj": {
749
+ "act_bits": 16,
750
+ "bits": 8,
751
+ "data_type": "int",
752
+ "group_size": 128,
753
+ "sym": true
754
+ },
755
+ "model.layers.14.mlp.router.classifier": {
756
+ "act_bits": 16,
757
+ "bits": 8,
758
+ "data_type": "int",
759
+ "group_size": 128,
760
+ "sym": true
761
+ },
762
+ "model.layers.14.mlps.0.down_proj": {
763
+ "act_bits": 16,
764
+ "bits": 8,
765
+ "data_type": "int",
766
+ "group_size": 128,
767
+ "sym": true
768
+ },
769
+ "model.layers.14.mlps.0.gate_proj": {
770
+ "act_bits": 16,
771
+ "bits": 8,
772
+ "data_type": "int",
773
+ "group_size": 128,
774
+ "sym": true
775
+ },
776
+ "model.layers.14.mlps.0.up_proj": {
777
+ "act_bits": 16,
778
+ "bits": 8,
779
+ "data_type": "int",
780
+ "group_size": 128,
781
+ "sym": true
782
+ },
783
+ "model.layers.14.mlps.1.down_proj": {
784
+ "act_bits": 16,
785
+ "bits": 8,
786
+ "data_type": "int",
787
+ "group_size": 128,
788
+ "sym": true
789
+ },
790
+ "model.layers.14.mlps.1.gate_proj": {
791
+ "act_bits": 16,
792
+ "bits": 8,
793
+ "data_type": "int",
794
+ "group_size": 128,
795
+ "sym": true
796
+ },
797
+ "model.layers.14.mlps.1.up_proj": {
798
+ "act_bits": 16,
799
+ "bits": 8,
800
+ "data_type": "int",
801
+ "group_size": 128,
802
+ "sym": true
803
+ },
804
+ "model.layers.14.self_attn.0.kv_a_proj_with_mqa": {
805
+ "act_bits": 16,
806
+ "bits": 8,
807
+ "data_type": "int",
808
+ "group_size": 128,
809
+ "sym": true
810
+ },
811
+ "model.layers.14.self_attn.0.kv_b_proj": {
812
+ "act_bits": 16,
813
+ "bits": 8,
814
+ "data_type": "int",
815
+ "group_size": 128,
816
+ "sym": true
817
+ },
818
+ "model.layers.14.self_attn.0.o_proj": {
819
+ "act_bits": 16,
820
+ "bits": 8,
821
+ "data_type": "int",
822
+ "group_size": 128,
823
+ "sym": true
824
+ },
825
+ "model.layers.14.self_attn.0.q_a_proj": {
826
+ "act_bits": 16,
827
+ "bits": 8,
828
+ "data_type": "int",
829
+ "group_size": 128,
830
+ "sym": true
831
+ },
832
+ "model.layers.14.self_attn.0.q_b_proj": {
833
+ "act_bits": 16,
834
+ "bits": 8,
835
+ "data_type": "int",
836
+ "group_size": 128,
837
+ "sym": true
838
+ },
839
+ "model.layers.14.self_attn.1.kv_a_proj_with_mqa": {
840
+ "act_bits": 16,
841
+ "bits": 8,
842
+ "data_type": "int",
843
+ "group_size": 128,
844
+ "sym": true
845
+ },
846
+ "model.layers.14.self_attn.1.kv_b_proj": {
847
+ "act_bits": 16,
848
+ "bits": 8,
849
+ "data_type": "int",
850
+ "group_size": 128,
851
+ "sym": true
852
+ },
853
+ "model.layers.14.self_attn.1.o_proj": {
854
+ "act_bits": 16,
855
+ "bits": 8,
856
+ "data_type": "int",
857
+ "group_size": 128,
858
+ "sym": true
859
+ },
860
+ "model.layers.14.self_attn.1.q_a_proj": {
861
+ "act_bits": 16,
862
+ "bits": 8,
863
+ "data_type": "int",
864
+ "group_size": 128,
865
+ "sym": true
866
+ },
867
+ "model.layers.14.self_attn.1.q_b_proj": {
868
+ "act_bits": 16,
869
+ "bits": 8,
870
+ "data_type": "int",
871
+ "group_size": 128,
872
+ "sym": true
873
+ },
874
+ "model.layers.15.mlp.router.classifier": {
875
+ "act_bits": 16,
876
+ "bits": 8,
877
+ "data_type": "int",
878
+ "group_size": 128,
879
+ "sym": true
880
+ },
881
+ "model.layers.15.mlps.0.down_proj": {
882
+ "act_bits": 16,
883
+ "bits": 8,
884
+ "data_type": "int",
885
+ "group_size": 128,
886
+ "sym": true
887
+ },
888
+ "model.layers.15.mlps.0.gate_proj": {
889
+ "act_bits": 16,
890
+ "bits": 8,
891
+ "data_type": "int",
892
+ "group_size": 128,
893
+ "sym": true
894
+ },
895
+ "model.layers.15.mlps.0.up_proj": {
896
+ "act_bits": 16,
897
+ "bits": 8,
898
+ "data_type": "int",
899
+ "group_size": 128,
900
+ "sym": true
901
+ },
902
+ "model.layers.15.mlps.1.down_proj": {
903
+ "act_bits": 16,
904
+ "bits": 8,
905
+ "data_type": "int",
906
+ "group_size": 128,
907
+ "sym": true
908
+ },
909
+ "model.layers.15.mlps.1.gate_proj": {
910
+ "act_bits": 16,
911
+ "bits": 8,
912
+ "data_type": "int",
913
+ "group_size": 128,
914
+ "sym": true
915
+ },
916
+ "model.layers.15.mlps.1.up_proj": {
917
+ "act_bits": 16,
918
+ "bits": 8,
919
+ "data_type": "int",
920
+ "group_size": 128,
921
+ "sym": true
922
+ },
923
+ "model.layers.15.self_attn.0.kv_a_proj_with_mqa": {
924
+ "act_bits": 16,
925
+ "bits": 8,
926
+ "data_type": "int",
927
+ "group_size": 128,
928
+ "sym": true
929
+ },
930
+ "model.layers.15.self_attn.0.kv_b_proj": {
931
+ "act_bits": 16,
932
+ "bits": 8,
933
+ "data_type": "int",
934
+ "group_size": 128,
935
+ "sym": true
936
+ },
937
+ "model.layers.15.self_attn.0.o_proj": {
938
+ "act_bits": 16,
939
+ "bits": 8,
940
+ "data_type": "int",
941
+ "group_size": 128,
942
+ "sym": true
943
+ },
944
+ "model.layers.15.self_attn.0.q_a_proj": {
945
+ "act_bits": 16,
946
+ "bits": 8,
947
+ "data_type": "int",
948
+ "group_size": 128,
949
+ "sym": true
950
+ },
951
+ "model.layers.15.self_attn.0.q_b_proj": {
952
+ "act_bits": 16,
953
+ "bits": 8,
954
+ "data_type": "int",
955
+ "group_size": 128,
956
+ "sym": true
957
+ },
958
+ "model.layers.15.self_attn.1.kv_a_proj_with_mqa": {
959
+ "act_bits": 16,
960
+ "bits": 8,
961
+ "data_type": "int",
962
+ "group_size": 128,
963
+ "sym": true
964
+ },
965
+ "model.layers.15.self_attn.1.kv_b_proj": {
966
+ "act_bits": 16,
967
+ "bits": 8,
968
+ "data_type": "int",
969
+ "group_size": 128,
970
+ "sym": true
971
+ },
972
+ "model.layers.15.self_attn.1.o_proj": {
973
+ "act_bits": 16,
974
+ "bits": 8,
975
+ "data_type": "int",
976
+ "group_size": 128,
977
+ "sym": true
978
+ },
979
+ "model.layers.15.self_attn.1.q_a_proj": {
980
+ "act_bits": 16,
981
+ "bits": 8,
982
+ "data_type": "int",
983
+ "group_size": 128,
984
+ "sym": true
985
+ },
986
+ "model.layers.15.self_attn.1.q_b_proj": {
987
+ "act_bits": 16,
988
+ "bits": 8,
989
+ "data_type": "int",
990
+ "group_size": 128,
991
+ "sym": true
992
+ },
993
+ "model.layers.16.mlp.router.classifier": {
994
+ "act_bits": 16,
995
+ "bits": 8,
996
+ "data_type": "int",
997
+ "group_size": 128,
998
+ "sym": true
999
+ },
1000
+ "model.layers.16.mlps.0.down_proj": {
1001
+ "act_bits": 16,
1002
+ "bits": 8,
1003
+ "data_type": "int",
1004
+ "group_size": 128,
1005
+ "sym": true
1006
+ },
1007
+ "model.layers.16.mlps.0.gate_proj": {
1008
+ "act_bits": 16,
1009
+ "bits": 8,
1010
+ "data_type": "int",
1011
+ "group_size": 128,
1012
+ "sym": true
1013
+ },
1014
+ "model.layers.16.mlps.0.up_proj": {
1015
+ "act_bits": 16,
1016
+ "bits": 8,
1017
+ "data_type": "int",
1018
+ "group_size": 128,
1019
+ "sym": true
1020
+ },
1021
+ "model.layers.16.mlps.1.down_proj": {
1022
+ "act_bits": 16,
1023
+ "bits": 8,
1024
+ "data_type": "int",
1025
+ "group_size": 128,
1026
+ "sym": true
1027
+ },
1028
+ "model.layers.16.mlps.1.gate_proj": {
1029
+ "act_bits": 16,
1030
+ "bits": 8,
1031
+ "data_type": "int",
1032
+ "group_size": 128,
1033
+ "sym": true
1034
+ },
1035
+ "model.layers.16.mlps.1.up_proj": {
1036
+ "act_bits": 16,
1037
+ "bits": 8,
1038
+ "data_type": "int",
1039
+ "group_size": 128,
1040
+ "sym": true
1041
+ },
1042
+ "model.layers.16.self_attn.0.kv_a_proj_with_mqa": {
1043
+ "act_bits": 16,
1044
+ "bits": 8,
1045
+ "data_type": "int",
1046
+ "group_size": 128,
1047
+ "sym": true
1048
+ },
1049
+ "model.layers.16.self_attn.0.kv_b_proj": {
1050
+ "act_bits": 16,
1051
+ "bits": 8,
1052
+ "data_type": "int",
1053
+ "group_size": 128,
1054
+ "sym": true
1055
+ },
1056
+ "model.layers.16.self_attn.0.o_proj": {
1057
+ "act_bits": 16,
1058
+ "bits": 8,
1059
+ "data_type": "int",
1060
+ "group_size": 128,
1061
+ "sym": true
1062
+ },
1063
+ "model.layers.16.self_attn.0.q_a_proj": {
1064
+ "act_bits": 16,
1065
+ "bits": 8,
1066
+ "data_type": "int",
1067
+ "group_size": 128,
1068
+ "sym": true
1069
+ },
1070
+ "model.layers.16.self_attn.0.q_b_proj": {
1071
+ "act_bits": 16,
1072
+ "bits": 8,
1073
+ "data_type": "int",
1074
+ "group_size": 128,
1075
+ "sym": true
1076
+ },
1077
+ "model.layers.16.self_attn.1.kv_a_proj_with_mqa": {
1078
+ "act_bits": 16,
1079
+ "bits": 8,
1080
+ "data_type": "int",
1081
+ "group_size": 128,
1082
+ "sym": true
1083
+ },
1084
+ "model.layers.16.self_attn.1.kv_b_proj": {
1085
+ "act_bits": 16,
1086
+ "bits": 8,
1087
+ "data_type": "int",
1088
+ "group_size": 128,
1089
+ "sym": true
1090
+ },
1091
+ "model.layers.16.self_attn.1.o_proj": {
1092
+ "act_bits": 16,
1093
+ "bits": 8,
1094
+ "data_type": "int",
1095
+ "group_size": 128,
1096
+ "sym": true
1097
+ },
1098
+ "model.layers.16.self_attn.1.q_a_proj": {
1099
+ "act_bits": 16,
1100
+ "bits": 8,
1101
+ "data_type": "int",
1102
+ "group_size": 128,
1103
+ "sym": true
1104
+ },
1105
+ "model.layers.16.self_attn.1.q_b_proj": {
1106
+ "act_bits": 16,
1107
+ "bits": 8,
1108
+ "data_type": "int",
1109
+ "group_size": 128,
1110
+ "sym": true
1111
+ },
1112
+ "model.layers.17.mlp.router.classifier": {
1113
+ "act_bits": 16,
1114
+ "bits": 8,
1115
+ "data_type": "int",
1116
+ "group_size": 128,
1117
+ "sym": true
1118
+ },
1119
+ "model.layers.17.mlps.0.down_proj": {
1120
+ "act_bits": 16,
1121
+ "bits": 8,
1122
+ "data_type": "int",
1123
+ "group_size": 128,
1124
+ "sym": true
1125
+ },
1126
+ "model.layers.17.mlps.0.gate_proj": {
1127
+ "act_bits": 16,
1128
+ "bits": 8,
1129
+ "data_type": "int",
1130
+ "group_size": 128,
1131
+ "sym": true
1132
+ },
1133
+ "model.layers.17.mlps.0.up_proj": {
1134
+ "act_bits": 16,
1135
+ "bits": 8,
1136
+ "data_type": "int",
1137
+ "group_size": 128,
1138
+ "sym": true
1139
+ },
1140
+ "model.layers.17.mlps.1.down_proj": {
1141
+ "act_bits": 16,
1142
+ "bits": 8,
1143
+ "data_type": "int",
1144
+ "group_size": 128,
1145
+ "sym": true
1146
+ },
1147
+ "model.layers.17.mlps.1.gate_proj": {
1148
+ "act_bits": 16,
1149
+ "bits": 8,
1150
+ "data_type": "int",
1151
+ "group_size": 128,
1152
+ "sym": true
1153
+ },
1154
+ "model.layers.17.mlps.1.up_proj": {
1155
+ "act_bits": 16,
1156
+ "bits": 8,
1157
+ "data_type": "int",
1158
+ "group_size": 128,
1159
+ "sym": true
1160
+ },
1161
+ "model.layers.17.self_attn.0.kv_a_proj_with_mqa": {
1162
+ "act_bits": 16,
1163
+ "bits": 8,
1164
+ "data_type": "int",
1165
+ "group_size": 128,
1166
+ "sym": true
1167
+ },
1168
+ "model.layers.17.self_attn.0.kv_b_proj": {
1169
+ "act_bits": 16,
1170
+ "bits": 8,
1171
+ "data_type": "int",
1172
+ "group_size": 128,
1173
+ "sym": true
1174
+ },
1175
+ "model.layers.17.self_attn.0.o_proj": {
1176
+ "act_bits": 16,
1177
+ "bits": 8,
1178
+ "data_type": "int",
1179
+ "group_size": 128,
1180
+ "sym": true
1181
+ },
1182
+ "model.layers.17.self_attn.0.q_a_proj": {
1183
+ "act_bits": 16,
1184
+ "bits": 8,
1185
+ "data_type": "int",
1186
+ "group_size": 128,
1187
+ "sym": true
1188
+ },
1189
+ "model.layers.17.self_attn.0.q_b_proj": {
1190
+ "act_bits": 16,
1191
+ "bits": 8,
1192
+ "data_type": "int",
1193
+ "group_size": 128,
1194
+ "sym": true
1195
+ },
1196
+ "model.layers.17.self_attn.1.kv_a_proj_with_mqa": {
1197
+ "act_bits": 16,
1198
+ "bits": 8,
1199
+ "data_type": "int",
1200
+ "group_size": 128,
1201
+ "sym": true
1202
+ },
1203
+ "model.layers.17.self_attn.1.kv_b_proj": {
1204
+ "act_bits": 16,
1205
+ "bits": 8,
1206
+ "data_type": "int",
1207
+ "group_size": 128,
1208
+ "sym": true
1209
+ },
1210
+ "model.layers.17.self_attn.1.o_proj": {
1211
+ "act_bits": 16,
1212
+ "bits": 8,
1213
+ "data_type": "int",
1214
+ "group_size": 128,
1215
+ "sym": true
1216
+ },
1217
+ "model.layers.17.self_attn.1.q_a_proj": {
1218
+ "act_bits": 16,
1219
+ "bits": 8,
1220
+ "data_type": "int",
1221
+ "group_size": 128,
1222
+ "sym": true
1223
+ },
1224
+ "model.layers.17.self_attn.1.q_b_proj": {
1225
+ "act_bits": 16,
1226
+ "bits": 8,
1227
+ "data_type": "int",
1228
+ "group_size": 128,
1229
+ "sym": true
1230
+ },
1231
+ "model.layers.18.mlp.router.classifier": {
1232
+ "act_bits": 16,
1233
+ "bits": 8,
1234
+ "data_type": "int",
1235
+ "group_size": 128,
1236
+ "sym": true
1237
+ },
1238
+ "model.layers.18.mlps.0.down_proj": {
1239
+ "act_bits": 16,
1240
+ "bits": 8,
1241
+ "data_type": "int",
1242
+ "group_size": 128,
1243
+ "sym": true
1244
+ },
1245
+ "model.layers.18.mlps.0.gate_proj": {
1246
+ "act_bits": 16,
1247
+ "bits": 8,
1248
+ "data_type": "int",
1249
+ "group_size": 128,
1250
+ "sym": true
1251
+ },
1252
+ "model.layers.18.mlps.0.up_proj": {
1253
+ "act_bits": 16,
1254
+ "bits": 8,
1255
+ "data_type": "int",
1256
+ "group_size": 128,
1257
+ "sym": true
1258
+ },
1259
+ "model.layers.18.mlps.1.down_proj": {
1260
+ "act_bits": 16,
1261
+ "bits": 8,
1262
+ "data_type": "int",
1263
+ "group_size": 128,
1264
+ "sym": true
1265
+ },
1266
+ "model.layers.18.mlps.1.gate_proj": {
1267
+ "act_bits": 16,
1268
+ "bits": 8,
1269
+ "data_type": "int",
1270
+ "group_size": 128,
1271
+ "sym": true
1272
+ },
1273
+ "model.layers.18.mlps.1.up_proj": {
1274
+ "act_bits": 16,
1275
+ "bits": 8,
1276
+ "data_type": "int",
1277
+ "group_size": 128,
1278
+ "sym": true
1279
+ },
1280
+ "model.layers.18.self_attn.0.kv_a_proj_with_mqa": {
1281
+ "act_bits": 16,
1282
+ "bits": 8,
1283
+ "data_type": "int",
1284
+ "group_size": 128,
1285
+ "sym": true
1286
+ },
1287
+ "model.layers.18.self_attn.0.kv_b_proj": {
1288
+ "act_bits": 16,
1289
+ "bits": 8,
1290
+ "data_type": "int",
1291
+ "group_size": 128,
1292
+ "sym": true
1293
+ },
1294
+ "model.layers.18.self_attn.0.o_proj": {
1295
+ "act_bits": 16,
1296
+ "bits": 8,
1297
+ "data_type": "int",
1298
+ "group_size": 128,
1299
+ "sym": true
1300
+ },
1301
+ "model.layers.18.self_attn.0.q_a_proj": {
1302
+ "act_bits": 16,
1303
+ "bits": 8,
1304
+ "data_type": "int",
1305
+ "group_size": 128,
1306
+ "sym": true
1307
+ },
1308
+ "model.layers.18.self_attn.0.q_b_proj": {
1309
+ "act_bits": 16,
1310
+ "bits": 8,
1311
+ "data_type": "int",
1312
+ "group_size": 128,
1313
+ "sym": true
1314
+ },
1315
+ "model.layers.18.self_attn.1.kv_a_proj_with_mqa": {
1316
+ "act_bits": 16,
1317
+ "bits": 8,
1318
+ "data_type": "int",
1319
+ "group_size": 128,
1320
+ "sym": true
1321
+ },
1322
+ "model.layers.18.self_attn.1.kv_b_proj": {
1323
+ "act_bits": 16,
1324
+ "bits": 8,
1325
+ "data_type": "int",
1326
+ "group_size": 128,
1327
+ "sym": true
1328
+ },
1329
+ "model.layers.18.self_attn.1.o_proj": {
1330
+ "act_bits": 16,
1331
+ "bits": 8,
1332
+ "data_type": "int",
1333
+ "group_size": 128,
1334
+ "sym": true
1335
+ },
1336
+ "model.layers.18.self_attn.1.q_a_proj": {
1337
+ "act_bits": 16,
1338
+ "bits": 8,
1339
+ "data_type": "int",
1340
+ "group_size": 128,
1341
+ "sym": true
1342
+ },
1343
+ "model.layers.18.self_attn.1.q_b_proj": {
1344
+ "act_bits": 16,
1345
+ "bits": 8,
1346
+ "data_type": "int",
1347
+ "group_size": 128,
1348
+ "sym": true
1349
+ },
1350
+ "model.layers.19.mlp.router.classifier": {
1351
+ "act_bits": 16,
1352
+ "bits": 8,
1353
+ "data_type": "int",
1354
+ "group_size": 128,
1355
+ "sym": true
1356
+ },
1357
+ "model.layers.19.mlps.0.down_proj": {
1358
+ "act_bits": 16,
1359
+ "bits": 8,
1360
+ "data_type": "int",
1361
+ "group_size": 128,
1362
+ "sym": true
1363
+ },
1364
+ "model.layers.19.mlps.0.gate_proj": {
1365
+ "act_bits": 16,
1366
+ "bits": 8,
1367
+ "data_type": "int",
1368
+ "group_size": 128,
1369
+ "sym": true
1370
+ },
1371
+ "model.layers.19.mlps.0.up_proj": {
1372
+ "act_bits": 16,
1373
+ "bits": 8,
1374
+ "data_type": "int",
1375
+ "group_size": 128,
1376
+ "sym": true
1377
+ },
1378
+ "model.layers.19.mlps.1.down_proj": {
1379
+ "act_bits": 16,
1380
+ "bits": 8,
1381
+ "data_type": "int",
1382
+ "group_size": 128,
1383
+ "sym": true
1384
+ },
1385
+ "model.layers.19.mlps.1.gate_proj": {
1386
+ "act_bits": 16,
1387
+ "bits": 8,
1388
+ "data_type": "int",
1389
+ "group_size": 128,
1390
+ "sym": true
1391
+ },
1392
+ "model.layers.19.mlps.1.up_proj": {
1393
+ "act_bits": 16,
1394
+ "bits": 8,
1395
+ "data_type": "int",
1396
+ "group_size": 128,
1397
+ "sym": true
1398
+ },
1399
+ "model.layers.19.self_attn.0.kv_a_proj_with_mqa": {
1400
+ "act_bits": 16,
1401
+ "bits": 8,
1402
+ "data_type": "int",
1403
+ "group_size": 128,
1404
+ "sym": true
1405
+ },
1406
+ "model.layers.19.self_attn.0.kv_b_proj": {
1407
+ "act_bits": 16,
1408
+ "bits": 8,
1409
+ "data_type": "int",
1410
+ "group_size": 128,
1411
+ "sym": true
1412
+ },
1413
+ "model.layers.19.self_attn.0.o_proj": {
1414
+ "act_bits": 16,
1415
+ "bits": 8,
1416
+ "data_type": "int",
1417
+ "group_size": 128,
1418
+ "sym": true
1419
+ },
1420
+ "model.layers.19.self_attn.0.q_a_proj": {
1421
+ "act_bits": 16,
1422
+ "bits": 8,
1423
+ "data_type": "int",
1424
+ "group_size": 128,
1425
+ "sym": true
1426
+ },
1427
+ "model.layers.19.self_attn.0.q_b_proj": {
1428
+ "act_bits": 16,
1429
+ "bits": 8,
1430
+ "data_type": "int",
1431
+ "group_size": 128,
1432
+ "sym": true
1433
+ },
1434
+ "model.layers.19.self_attn.1.kv_a_proj_with_mqa": {
1435
+ "act_bits": 16,
1436
+ "bits": 8,
1437
+ "data_type": "int",
1438
+ "group_size": 128,
1439
+ "sym": true
1440
+ },
1441
+ "model.layers.19.self_attn.1.kv_b_proj": {
1442
+ "act_bits": 16,
1443
+ "bits": 8,
1444
+ "data_type": "int",
1445
+ "group_size": 128,
1446
+ "sym": true
1447
+ },
1448
+ "model.layers.19.self_attn.1.o_proj": {
1449
+ "act_bits": 16,
1450
+ "bits": 8,
1451
+ "data_type": "int",
1452
+ "group_size": 128,
1453
+ "sym": true
1454
+ },
1455
+ "model.layers.19.self_attn.1.q_a_proj": {
1456
+ "act_bits": 16,
1457
+ "bits": 8,
1458
+ "data_type": "int",
1459
+ "group_size": 128,
1460
+ "sym": true
1461
+ },
1462
+ "model.layers.19.self_attn.1.q_b_proj": {
1463
+ "act_bits": 16,
1464
+ "bits": 8,
1465
+ "data_type": "int",
1466
+ "group_size": 128,
1467
+ "sym": true
1468
+ },
1469
+ "model.layers.2.mlp.router.classifier": {
1470
+ "act_bits": 16,
1471
+ "bits": 8,
1472
+ "data_type": "int",
1473
+ "group_size": 128,
1474
+ "sym": true
1475
+ },
1476
+ "model.layers.2.mlps.0.down_proj": {
1477
+ "act_bits": 16,
1478
+ "bits": 8,
1479
+ "data_type": "int",
1480
+ "group_size": 128,
1481
+ "sym": true
1482
+ },
1483
+ "model.layers.2.mlps.0.gate_proj": {
1484
+ "act_bits": 16,
1485
+ "bits": 8,
1486
+ "data_type": "int",
1487
+ "group_size": 128,
1488
+ "sym": true
1489
+ },
1490
+ "model.layers.2.mlps.0.up_proj": {
1491
+ "act_bits": 16,
1492
+ "bits": 8,
1493
+ "data_type": "int",
1494
+ "group_size": 128,
1495
+ "sym": true
1496
+ },
1497
+ "model.layers.2.mlps.1.down_proj": {
1498
+ "act_bits": 16,
1499
+ "bits": 8,
1500
+ "data_type": "int",
1501
+ "group_size": 128,
1502
+ "sym": true
1503
+ },
1504
+ "model.layers.2.mlps.1.gate_proj": {
1505
+ "act_bits": 16,
1506
+ "bits": 8,
1507
+ "data_type": "int",
1508
+ "group_size": 128,
1509
+ "sym": true
1510
+ },
1511
+ "model.layers.2.mlps.1.up_proj": {
1512
+ "act_bits": 16,
1513
+ "bits": 8,
1514
+ "data_type": "int",
1515
+ "group_size": 128,
1516
+ "sym": true
1517
+ },
1518
+ "model.layers.2.self_attn.0.kv_a_proj_with_mqa": {
1519
+ "act_bits": 16,
1520
+ "bits": 8,
1521
+ "data_type": "int",
1522
+ "group_size": 128,
1523
+ "sym": true
1524
+ },
1525
+ "model.layers.2.self_attn.0.kv_b_proj": {
1526
+ "act_bits": 16,
1527
+ "bits": 8,
1528
+ "data_type": "int",
1529
+ "group_size": 128,
1530
+ "sym": true
1531
+ },
1532
+ "model.layers.2.self_attn.0.o_proj": {
1533
+ "act_bits": 16,
1534
+ "bits": 8,
1535
+ "data_type": "int",
1536
+ "group_size": 128,
1537
+ "sym": true
1538
+ },
1539
+ "model.layers.2.self_attn.0.q_a_proj": {
1540
+ "act_bits": 16,
1541
+ "bits": 8,
1542
+ "data_type": "int",
1543
+ "group_size": 128,
1544
+ "sym": true
1545
+ },
1546
+ "model.layers.2.self_attn.0.q_b_proj": {
1547
+ "act_bits": 16,
1548
+ "bits": 8,
1549
+ "data_type": "int",
1550
+ "group_size": 128,
1551
+ "sym": true
1552
+ },
1553
+ "model.layers.2.self_attn.1.kv_a_proj_with_mqa": {
1554
+ "act_bits": 16,
1555
+ "bits": 8,
1556
+ "data_type": "int",
1557
+ "group_size": 128,
1558
+ "sym": true
1559
+ },
1560
+ "model.layers.2.self_attn.1.kv_b_proj": {
1561
+ "act_bits": 16,
1562
+ "bits": 8,
1563
+ "data_type": "int",
1564
+ "group_size": 128,
1565
+ "sym": true
1566
+ },
1567
+ "model.layers.2.self_attn.1.o_proj": {
1568
+ "act_bits": 16,
1569
+ "bits": 8,
1570
+ "data_type": "int",
1571
+ "group_size": 128,
1572
+ "sym": true
1573
+ },
1574
+ "model.layers.2.self_attn.1.q_a_proj": {
1575
+ "act_bits": 16,
1576
+ "bits": 8,
1577
+ "data_type": "int",
1578
+ "group_size": 128,
1579
+ "sym": true
1580
+ },
1581
+ "model.layers.2.self_attn.1.q_b_proj": {
1582
+ "act_bits": 16,
1583
+ "bits": 8,
1584
+ "data_type": "int",
1585
+ "group_size": 128,
1586
+ "sym": true
1587
+ },
1588
+ "model.layers.20.mlp.router.classifier": {
1589
+ "act_bits": 16,
1590
+ "bits": 8,
1591
+ "data_type": "int",
1592
+ "group_size": 128,
1593
+ "sym": true
1594
+ },
1595
+ "model.layers.20.mlps.0.down_proj": {
1596
+ "act_bits": 16,
1597
+ "bits": 8,
1598
+ "data_type": "int",
1599
+ "group_size": 128,
1600
+ "sym": true
1601
+ },
1602
+ "model.layers.20.mlps.0.gate_proj": {
1603
+ "act_bits": 16,
1604
+ "bits": 8,
1605
+ "data_type": "int",
1606
+ "group_size": 128,
1607
+ "sym": true
1608
+ },
1609
+ "model.layers.20.mlps.0.up_proj": {
1610
+ "act_bits": 16,
1611
+ "bits": 8,
1612
+ "data_type": "int",
1613
+ "group_size": 128,
1614
+ "sym": true
1615
+ },
1616
+ "model.layers.20.mlps.1.down_proj": {
1617
+ "act_bits": 16,
1618
+ "bits": 8,
1619
+ "data_type": "int",
1620
+ "group_size": 128,
1621
+ "sym": true
1622
+ },
1623
+ "model.layers.20.mlps.1.gate_proj": {
1624
+ "act_bits": 16,
1625
+ "bits": 8,
1626
+ "data_type": "int",
1627
+ "group_size": 128,
1628
+ "sym": true
1629
+ },
1630
+ "model.layers.20.mlps.1.up_proj": {
1631
+ "act_bits": 16,
1632
+ "bits": 8,
1633
+ "data_type": "int",
1634
+ "group_size": 128,
1635
+ "sym": true
1636
+ },
1637
+ "model.layers.20.self_attn.0.kv_a_proj_with_mqa": {
1638
+ "act_bits": 16,
1639
+ "bits": 8,
1640
+ "data_type": "int",
1641
+ "group_size": 128,
1642
+ "sym": true
1643
+ },
1644
+ "model.layers.20.self_attn.0.kv_b_proj": {
1645
+ "act_bits": 16,
1646
+ "bits": 8,
1647
+ "data_type": "int",
1648
+ "group_size": 128,
1649
+ "sym": true
1650
+ },
1651
+ "model.layers.20.self_attn.0.o_proj": {
1652
+ "act_bits": 16,
1653
+ "bits": 8,
1654
+ "data_type": "int",
1655
+ "group_size": 128,
1656
+ "sym": true
1657
+ },
1658
+ "model.layers.20.self_attn.0.q_a_proj": {
1659
+ "act_bits": 16,
1660
+ "bits": 8,
1661
+ "data_type": "int",
1662
+ "group_size": 128,
1663
+ "sym": true
1664
+ },
1665
+ "model.layers.20.self_attn.0.q_b_proj": {
1666
+ "act_bits": 16,
1667
+ "bits": 8,
1668
+ "data_type": "int",
1669
+ "group_size": 128,
1670
+ "sym": true
1671
+ },
1672
+ "model.layers.20.self_attn.1.kv_a_proj_with_mqa": {
1673
+ "act_bits": 16,
1674
+ "bits": 8,
1675
+ "data_type": "int",
1676
+ "group_size": 128,
1677
+ "sym": true
1678
+ },
1679
+ "model.layers.20.self_attn.1.kv_b_proj": {
1680
+ "act_bits": 16,
1681
+ "bits": 8,
1682
+ "data_type": "int",
1683
+ "group_size": 128,
1684
+ "sym": true
1685
+ },
1686
+ "model.layers.20.self_attn.1.o_proj": {
1687
+ "act_bits": 16,
1688
+ "bits": 8,
1689
+ "data_type": "int",
1690
+ "group_size": 128,
1691
+ "sym": true
1692
+ },
1693
+ "model.layers.20.self_attn.1.q_a_proj": {
1694
+ "act_bits": 16,
1695
+ "bits": 8,
1696
+ "data_type": "int",
1697
+ "group_size": 128,
1698
+ "sym": true
1699
+ },
1700
+ "model.layers.20.self_attn.1.q_b_proj": {
1701
+ "act_bits": 16,
1702
+ "bits": 8,
1703
+ "data_type": "int",
1704
+ "group_size": 128,
1705
+ "sym": true
1706
+ },
1707
+ "model.layers.21.mlp.router.classifier": {
1708
+ "act_bits": 16,
1709
+ "bits": 8,
1710
+ "data_type": "int",
1711
+ "group_size": 128,
1712
+ "sym": true
1713
+ },
1714
+ "model.layers.21.mlps.0.down_proj": {
1715
+ "act_bits": 16,
1716
+ "bits": 8,
1717
+ "data_type": "int",
1718
+ "group_size": 128,
1719
+ "sym": true
1720
+ },
1721
+ "model.layers.21.mlps.0.gate_proj": {
1722
+ "act_bits": 16,
1723
+ "bits": 8,
1724
+ "data_type": "int",
1725
+ "group_size": 128,
1726
+ "sym": true
1727
+ },
1728
+ "model.layers.21.mlps.0.up_proj": {
1729
+ "act_bits": 16,
1730
+ "bits": 8,
1731
+ "data_type": "int",
1732
+ "group_size": 128,
1733
+ "sym": true
1734
+ },
1735
+ "model.layers.21.mlps.1.down_proj": {
1736
+ "act_bits": 16,
1737
+ "bits": 8,
1738
+ "data_type": "int",
1739
+ "group_size": 128,
1740
+ "sym": true
1741
+ },
1742
+ "model.layers.21.mlps.1.gate_proj": {
1743
+ "act_bits": 16,
1744
+ "bits": 8,
1745
+ "data_type": "int",
1746
+ "group_size": 128,
1747
+ "sym": true
1748
+ },
1749
+ "model.layers.21.mlps.1.up_proj": {
1750
+ "act_bits": 16,
1751
+ "bits": 8,
1752
+ "data_type": "int",
1753
+ "group_size": 128,
1754
+ "sym": true
1755
+ },
1756
+ "model.layers.21.self_attn.0.kv_a_proj_with_mqa": {
1757
+ "act_bits": 16,
1758
+ "bits": 8,
1759
+ "data_type": "int",
1760
+ "group_size": 128,
1761
+ "sym": true
1762
+ },
1763
+ "model.layers.21.self_attn.0.kv_b_proj": {
1764
+ "act_bits": 16,
1765
+ "bits": 8,
1766
+ "data_type": "int",
1767
+ "group_size": 128,
1768
+ "sym": true
1769
+ },
1770
+ "model.layers.21.self_attn.0.o_proj": {
1771
+ "act_bits": 16,
1772
+ "bits": 8,
1773
+ "data_type": "int",
1774
+ "group_size": 128,
1775
+ "sym": true
1776
+ },
1777
+ "model.layers.21.self_attn.0.q_a_proj": {
1778
+ "act_bits": 16,
1779
+ "bits": 8,
1780
+ "data_type": "int",
1781
+ "group_size": 128,
1782
+ "sym": true
1783
+ },
1784
+ "model.layers.21.self_attn.0.q_b_proj": {
1785
+ "act_bits": 16,
1786
+ "bits": 8,
1787
+ "data_type": "int",
1788
+ "group_size": 128,
1789
+ "sym": true
1790
+ },
1791
+ "model.layers.21.self_attn.1.kv_a_proj_with_mqa": {
1792
+ "act_bits": 16,
1793
+ "bits": 8,
1794
+ "data_type": "int",
1795
+ "group_size": 128,
1796
+ "sym": true
1797
+ },
1798
+ "model.layers.21.self_attn.1.kv_b_proj": {
1799
+ "act_bits": 16,
1800
+ "bits": 8,
1801
+ "data_type": "int",
1802
+ "group_size": 128,
1803
+ "sym": true
1804
+ },
1805
+ "model.layers.21.self_attn.1.o_proj": {
1806
+ "act_bits": 16,
1807
+ "bits": 8,
1808
+ "data_type": "int",
1809
+ "group_size": 128,
1810
+ "sym": true
1811
+ },
1812
+ "model.layers.21.self_attn.1.q_a_proj": {
1813
+ "act_bits": 16,
1814
+ "bits": 8,
1815
+ "data_type": "int",
1816
+ "group_size": 128,
1817
+ "sym": true
1818
+ },
1819
+ "model.layers.21.self_attn.1.q_b_proj": {
1820
+ "act_bits": 16,
1821
+ "bits": 8,
1822
+ "data_type": "int",
1823
+ "group_size": 128,
1824
+ "sym": true
1825
+ },
1826
+ "model.layers.22.mlp.router.classifier": {
1827
+ "act_bits": 16,
1828
+ "bits": 8,
1829
+ "data_type": "int",
1830
+ "group_size": 128,
1831
+ "sym": true
1832
+ },
1833
+ "model.layers.22.mlps.0.down_proj": {
1834
+ "act_bits": 16,
1835
+ "bits": 8,
1836
+ "data_type": "int",
1837
+ "group_size": 128,
1838
+ "sym": true
1839
+ },
1840
+ "model.layers.22.mlps.0.gate_proj": {
1841
+ "act_bits": 16,
1842
+ "bits": 8,
1843
+ "data_type": "int",
1844
+ "group_size": 128,
1845
+ "sym": true
1846
+ },
1847
+ "model.layers.22.mlps.0.up_proj": {
1848
+ "act_bits": 16,
1849
+ "bits": 8,
1850
+ "data_type": "int",
1851
+ "group_size": 128,
1852
+ "sym": true
1853
+ },
1854
+ "model.layers.22.mlps.1.down_proj": {
1855
+ "act_bits": 16,
1856
+ "bits": 8,
1857
+ "data_type": "int",
1858
+ "group_size": 128,
1859
+ "sym": true
1860
+ },
1861
+ "model.layers.22.mlps.1.gate_proj": {
1862
+ "act_bits": 16,
1863
+ "bits": 8,
1864
+ "data_type": "int",
1865
+ "group_size": 128,
1866
+ "sym": true
1867
+ },
1868
+ "model.layers.22.mlps.1.up_proj": {
1869
+ "act_bits": 16,
1870
+ "bits": 8,
1871
+ "data_type": "int",
1872
+ "group_size": 128,
1873
+ "sym": true
1874
+ },
1875
+ "model.layers.22.self_attn.0.kv_a_proj_with_mqa": {
1876
+ "act_bits": 16,
1877
+ "bits": 8,
1878
+ "data_type": "int",
1879
+ "group_size": 128,
1880
+ "sym": true
1881
+ },
1882
+ "model.layers.22.self_attn.0.kv_b_proj": {
1883
+ "act_bits": 16,
1884
+ "bits": 8,
1885
+ "data_type": "int",
1886
+ "group_size": 128,
1887
+ "sym": true
1888
+ },
1889
+ "model.layers.22.self_attn.0.o_proj": {
1890
+ "act_bits": 16,
1891
+ "bits": 8,
1892
+ "data_type": "int",
1893
+ "group_size": 128,
1894
+ "sym": true
1895
+ },
1896
+ "model.layers.22.self_attn.0.q_a_proj": {
1897
+ "act_bits": 16,
1898
+ "bits": 8,
1899
+ "data_type": "int",
1900
+ "group_size": 128,
1901
+ "sym": true
1902
+ },
1903
+ "model.layers.22.self_attn.0.q_b_proj": {
1904
+ "act_bits": 16,
1905
+ "bits": 8,
1906
+ "data_type": "int",
1907
+ "group_size": 128,
1908
+ "sym": true
1909
+ },
1910
+ "model.layers.22.self_attn.1.kv_a_proj_with_mqa": {
1911
+ "act_bits": 16,
1912
+ "bits": 8,
1913
+ "data_type": "int",
1914
+ "group_size": 128,
1915
+ "sym": true
1916
+ },
1917
+ "model.layers.22.self_attn.1.kv_b_proj": {
1918
+ "act_bits": 16,
1919
+ "bits": 8,
1920
+ "data_type": "int",
1921
+ "group_size": 128,
1922
+ "sym": true
1923
+ },
1924
+ "model.layers.22.self_attn.1.o_proj": {
1925
+ "act_bits": 16,
1926
+ "bits": 8,
1927
+ "data_type": "int",
1928
+ "group_size": 128,
1929
+ "sym": true
1930
+ },
1931
+ "model.layers.22.self_attn.1.q_a_proj": {
1932
+ "act_bits": 16,
1933
+ "bits": 8,
1934
+ "data_type": "int",
1935
+ "group_size": 128,
1936
+ "sym": true
1937
+ },
1938
+ "model.layers.22.self_attn.1.q_b_proj": {
1939
+ "act_bits": 16,
1940
+ "bits": 8,
1941
+ "data_type": "int",
1942
+ "group_size": 128,
1943
+ "sym": true
1944
+ },
1945
+ "model.layers.23.mlp.router.classifier": {
1946
+ "act_bits": 16,
1947
+ "bits": 8,
1948
+ "data_type": "int",
1949
+ "group_size": 128,
1950
+ "sym": true
1951
+ },
1952
+ "model.layers.23.mlps.0.down_proj": {
1953
+ "act_bits": 16,
1954
+ "bits": 8,
1955
+ "data_type": "int",
1956
+ "group_size": 128,
1957
+ "sym": true
1958
+ },
1959
+ "model.layers.23.mlps.0.gate_proj": {
1960
+ "act_bits": 16,
1961
+ "bits": 8,
1962
+ "data_type": "int",
1963
+ "group_size": 128,
1964
+ "sym": true
1965
+ },
1966
+ "model.layers.23.mlps.0.up_proj": {
1967
+ "act_bits": 16,
1968
+ "bits": 8,
1969
+ "data_type": "int",
1970
+ "group_size": 128,
1971
+ "sym": true
1972
+ },
1973
+ "model.layers.23.mlps.1.down_proj": {
1974
+ "act_bits": 16,
1975
+ "bits": 8,
1976
+ "data_type": "int",
1977
+ "group_size": 128,
1978
+ "sym": true
1979
+ },
1980
+ "model.layers.23.mlps.1.gate_proj": {
1981
+ "act_bits": 16,
1982
+ "bits": 8,
1983
+ "data_type": "int",
1984
+ "group_size": 128,
1985
+ "sym": true
1986
+ },
1987
+ "model.layers.23.mlps.1.up_proj": {
1988
+ "act_bits": 16,
1989
+ "bits": 8,
1990
+ "data_type": "int",
1991
+ "group_size": 128,
1992
+ "sym": true
1993
+ },
1994
+ "model.layers.23.self_attn.0.kv_a_proj_with_mqa": {
1995
+ "act_bits": 16,
1996
+ "bits": 8,
1997
+ "data_type": "int",
1998
+ "group_size": 128,
1999
+ "sym": true
2000
+ },
2001
+ "model.layers.23.self_attn.0.kv_b_proj": {
2002
+ "act_bits": 16,
2003
+ "bits": 8,
2004
+ "data_type": "int",
2005
+ "group_size": 128,
2006
+ "sym": true
2007
+ },
2008
+ "model.layers.23.self_attn.0.o_proj": {
2009
+ "act_bits": 16,
2010
+ "bits": 8,
2011
+ "data_type": "int",
2012
+ "group_size": 128,
2013
+ "sym": true
2014
+ },
2015
+ "model.layers.23.self_attn.0.q_a_proj": {
2016
+ "act_bits": 16,
2017
+ "bits": 8,
2018
+ "data_type": "int",
2019
+ "group_size": 128,
2020
+ "sym": true
2021
+ },
2022
+ "model.layers.23.self_attn.0.q_b_proj": {
2023
+ "act_bits": 16,
2024
+ "bits": 8,
2025
+ "data_type": "int",
2026
+ "group_size": 128,
2027
+ "sym": true
2028
+ },
2029
+ "model.layers.23.self_attn.1.kv_a_proj_with_mqa": {
2030
+ "act_bits": 16,
2031
+ "bits": 8,
2032
+ "data_type": "int",
2033
+ "group_size": 128,
2034
+ "sym": true
2035
+ },
2036
+ "model.layers.23.self_attn.1.kv_b_proj": {
2037
+ "act_bits": 16,
2038
+ "bits": 8,
2039
+ "data_type": "int",
2040
+ "group_size": 128,
2041
+ "sym": true
2042
+ },
2043
+ "model.layers.23.self_attn.1.o_proj": {
2044
+ "act_bits": 16,
2045
+ "bits": 8,
2046
+ "data_type": "int",
2047
+ "group_size": 128,
2048
+ "sym": true
2049
+ },
2050
+ "model.layers.23.self_attn.1.q_a_proj": {
2051
+ "act_bits": 16,
2052
+ "bits": 8,
2053
+ "data_type": "int",
2054
+ "group_size": 128,
2055
+ "sym": true
2056
+ },
2057
+ "model.layers.23.self_attn.1.q_b_proj": {
2058
+ "act_bits": 16,
2059
+ "bits": 8,
2060
+ "data_type": "int",
2061
+ "group_size": 128,
2062
+ "sym": true
2063
+ },
2064
+ "model.layers.24.mlp.router.classifier": {
2065
+ "act_bits": 16,
2066
+ "bits": 8,
2067
+ "data_type": "int",
2068
+ "group_size": 128,
2069
+ "sym": true
2070
+ },
2071
+ "model.layers.24.mlps.0.down_proj": {
2072
+ "act_bits": 16,
2073
+ "bits": 8,
2074
+ "data_type": "int",
2075
+ "group_size": 128,
2076
+ "sym": true
2077
+ },
2078
+ "model.layers.24.mlps.0.gate_proj": {
2079
+ "act_bits": 16,
2080
+ "bits": 8,
2081
+ "data_type": "int",
2082
+ "group_size": 128,
2083
+ "sym": true
2084
+ },
2085
+ "model.layers.24.mlps.0.up_proj": {
2086
+ "act_bits": 16,
2087
+ "bits": 8,
2088
+ "data_type": "int",
2089
+ "group_size": 128,
2090
+ "sym": true
2091
+ },
2092
+ "model.layers.24.mlps.1.down_proj": {
2093
+ "act_bits": 16,
2094
+ "bits": 8,
2095
+ "data_type": "int",
2096
+ "group_size": 128,
2097
+ "sym": true
2098
+ },
2099
+ "model.layers.24.mlps.1.gate_proj": {
2100
+ "act_bits": 16,
2101
+ "bits": 8,
2102
+ "data_type": "int",
2103
+ "group_size": 128,
2104
+ "sym": true
2105
+ },
2106
+ "model.layers.24.mlps.1.up_proj": {
2107
+ "act_bits": 16,
2108
+ "bits": 8,
2109
+ "data_type": "int",
2110
+ "group_size": 128,
2111
+ "sym": true
2112
+ },
2113
+ "model.layers.24.self_attn.0.kv_a_proj_with_mqa": {
2114
+ "act_bits": 16,
2115
+ "bits": 8,
2116
+ "data_type": "int",
2117
+ "group_size": 128,
2118
+ "sym": true
2119
+ },
2120
+ "model.layers.24.self_attn.0.kv_b_proj": {
2121
+ "act_bits": 16,
2122
+ "bits": 8,
2123
+ "data_type": "int",
2124
+ "group_size": 128,
2125
+ "sym": true
2126
+ },
2127
+ "model.layers.24.self_attn.0.o_proj": {
2128
+ "act_bits": 16,
2129
+ "bits": 8,
2130
+ "data_type": "int",
2131
+ "group_size": 128,
2132
+ "sym": true
2133
+ },
2134
+ "model.layers.24.self_attn.0.q_a_proj": {
2135
+ "act_bits": 16,
2136
+ "bits": 8,
2137
+ "data_type": "int",
2138
+ "group_size": 128,
2139
+ "sym": true
2140
+ },
2141
+ "model.layers.24.self_attn.0.q_b_proj": {
2142
+ "act_bits": 16,
2143
+ "bits": 8,
2144
+ "data_type": "int",
2145
+ "group_size": 128,
2146
+ "sym": true
2147
+ },
2148
+ "model.layers.24.self_attn.1.kv_a_proj_with_mqa": {
2149
+ "act_bits": 16,
2150
+ "bits": 8,
2151
+ "data_type": "int",
2152
+ "group_size": 128,
2153
+ "sym": true
2154
+ },
2155
+ "model.layers.24.self_attn.1.kv_b_proj": {
2156
+ "act_bits": 16,
2157
+ "bits": 8,
2158
+ "data_type": "int",
2159
+ "group_size": 128,
2160
+ "sym": true
2161
+ },
2162
+ "model.layers.24.self_attn.1.o_proj": {
2163
+ "act_bits": 16,
2164
+ "bits": 8,
2165
+ "data_type": "int",
2166
+ "group_size": 128,
2167
+ "sym": true
2168
+ },
2169
+ "model.layers.24.self_attn.1.q_a_proj": {
2170
+ "act_bits": 16,
2171
+ "bits": 8,
2172
+ "data_type": "int",
2173
+ "group_size": 128,
2174
+ "sym": true
2175
+ },
2176
+ "model.layers.24.self_attn.1.q_b_proj": {
2177
+ "act_bits": 16,
2178
+ "bits": 8,
2179
+ "data_type": "int",
2180
+ "group_size": 128,
2181
+ "sym": true
2182
+ },
2183
+ "model.layers.25.mlp.router.classifier": {
2184
+ "act_bits": 16,
2185
+ "bits": 8,
2186
+ "data_type": "int",
2187
+ "group_size": 128,
2188
+ "sym": true
2189
+ },
2190
+ "model.layers.25.mlps.0.down_proj": {
2191
+ "act_bits": 16,
2192
+ "bits": 8,
2193
+ "data_type": "int",
2194
+ "group_size": 128,
2195
+ "sym": true
2196
+ },
2197
+ "model.layers.25.mlps.0.gate_proj": {
2198
+ "act_bits": 16,
2199
+ "bits": 8,
2200
+ "data_type": "int",
2201
+ "group_size": 128,
2202
+ "sym": true
2203
+ },
2204
+ "model.layers.25.mlps.0.up_proj": {
2205
+ "act_bits": 16,
2206
+ "bits": 8,
2207
+ "data_type": "int",
2208
+ "group_size": 128,
2209
+ "sym": true
2210
+ },
2211
+ "model.layers.25.mlps.1.down_proj": {
2212
+ "act_bits": 16,
2213
+ "bits": 8,
2214
+ "data_type": "int",
2215
+ "group_size": 128,
2216
+ "sym": true
2217
+ },
2218
+ "model.layers.25.mlps.1.gate_proj": {
2219
+ "act_bits": 16,
2220
+ "bits": 8,
2221
+ "data_type": "int",
2222
+ "group_size": 128,
2223
+ "sym": true
2224
+ },
2225
+ "model.layers.25.mlps.1.up_proj": {
2226
+ "act_bits": 16,
2227
+ "bits": 8,
2228
+ "data_type": "int",
2229
+ "group_size": 128,
2230
+ "sym": true
2231
+ },
2232
+ "model.layers.25.self_attn.0.kv_a_proj_with_mqa": {
2233
+ "act_bits": 16,
2234
+ "bits": 8,
2235
+ "data_type": "int",
2236
+ "group_size": 128,
2237
+ "sym": true
2238
+ },
2239
+ "model.layers.25.self_attn.0.kv_b_proj": {
2240
+ "act_bits": 16,
2241
+ "bits": 8,
2242
+ "data_type": "int",
2243
+ "group_size": 128,
2244
+ "sym": true
2245
+ },
2246
+ "model.layers.25.self_attn.0.o_proj": {
2247
+ "act_bits": 16,
2248
+ "bits": 8,
2249
+ "data_type": "int",
2250
+ "group_size": 128,
2251
+ "sym": true
2252
+ },
2253
+ "model.layers.25.self_attn.0.q_a_proj": {
2254
+ "act_bits": 16,
2255
+ "bits": 8,
2256
+ "data_type": "int",
2257
+ "group_size": 128,
2258
+ "sym": true
2259
+ },
2260
+ "model.layers.25.self_attn.0.q_b_proj": {
2261
+ "act_bits": 16,
2262
+ "bits": 8,
2263
+ "data_type": "int",
2264
+ "group_size": 128,
2265
+ "sym": true
2266
+ },
2267
+ "model.layers.25.self_attn.1.kv_a_proj_with_mqa": {
2268
+ "act_bits": 16,
2269
+ "bits": 8,
2270
+ "data_type": "int",
2271
+ "group_size": 128,
2272
+ "sym": true
2273
+ },
2274
+ "model.layers.25.self_attn.1.kv_b_proj": {
2275
+ "act_bits": 16,
2276
+ "bits": 8,
2277
+ "data_type": "int",
2278
+ "group_size": 128,
2279
+ "sym": true
2280
+ },
2281
+ "model.layers.25.self_attn.1.o_proj": {
2282
+ "act_bits": 16,
2283
+ "bits": 8,
2284
+ "data_type": "int",
2285
+ "group_size": 128,
2286
+ "sym": true
2287
+ },
2288
+ "model.layers.25.self_attn.1.q_a_proj": {
2289
+ "act_bits": 16,
2290
+ "bits": 8,
2291
+ "data_type": "int",
2292
+ "group_size": 128,
2293
+ "sym": true
2294
+ },
2295
+ "model.layers.25.self_attn.1.q_b_proj": {
2296
+ "act_bits": 16,
2297
+ "bits": 8,
2298
+ "data_type": "int",
2299
+ "group_size": 128,
2300
+ "sym": true
2301
+ },
2302
+ "model.layers.26.mlp.router.classifier": {
2303
+ "act_bits": 16,
2304
+ "bits": 8,
2305
+ "data_type": "int",
2306
+ "group_size": 128,
2307
+ "sym": true
2308
+ },
2309
+ "model.layers.26.mlps.0.down_proj": {
2310
+ "act_bits": 16,
2311
+ "bits": 8,
2312
+ "data_type": "int",
2313
+ "group_size": 128,
2314
+ "sym": true
2315
+ },
2316
+ "model.layers.26.mlps.0.gate_proj": {
2317
+ "act_bits": 16,
2318
+ "bits": 8,
2319
+ "data_type": "int",
2320
+ "group_size": 128,
2321
+ "sym": true
2322
+ },
2323
+ "model.layers.26.mlps.0.up_proj": {
2324
+ "act_bits": 16,
2325
+ "bits": 8,
2326
+ "data_type": "int",
2327
+ "group_size": 128,
2328
+ "sym": true
2329
+ },
2330
+ "model.layers.26.mlps.1.down_proj": {
2331
+ "act_bits": 16,
2332
+ "bits": 8,
2333
+ "data_type": "int",
2334
+ "group_size": 128,
2335
+ "sym": true
2336
+ },
2337
+ "model.layers.26.mlps.1.gate_proj": {
2338
+ "act_bits": 16,
2339
+ "bits": 8,
2340
+ "data_type": "int",
2341
+ "group_size": 128,
2342
+ "sym": true
2343
+ },
2344
+ "model.layers.26.mlps.1.up_proj": {
2345
+ "act_bits": 16,
2346
+ "bits": 8,
2347
+ "data_type": "int",
2348
+ "group_size": 128,
2349
+ "sym": true
2350
+ },
2351
+ "model.layers.26.self_attn.0.kv_a_proj_with_mqa": {
2352
+ "act_bits": 16,
2353
+ "bits": 8,
2354
+ "data_type": "int",
2355
+ "group_size": 128,
2356
+ "sym": true
2357
+ },
2358
+ "model.layers.26.self_attn.0.kv_b_proj": {
2359
+ "act_bits": 16,
2360
+ "bits": 8,
2361
+ "data_type": "int",
2362
+ "group_size": 128,
2363
+ "sym": true
2364
+ },
2365
+ "model.layers.26.self_attn.0.o_proj": {
2366
+ "act_bits": 16,
2367
+ "bits": 8,
2368
+ "data_type": "int",
2369
+ "group_size": 128,
2370
+ "sym": true
2371
+ },
2372
+ "model.layers.26.self_attn.0.q_a_proj": {
2373
+ "act_bits": 16,
2374
+ "bits": 8,
2375
+ "data_type": "int",
2376
+ "group_size": 128,
2377
+ "sym": true
2378
+ },
2379
+ "model.layers.26.self_attn.0.q_b_proj": {
2380
+ "act_bits": 16,
2381
+ "bits": 8,
2382
+ "data_type": "int",
2383
+ "group_size": 128,
2384
+ "sym": true
2385
+ },
2386
+ "model.layers.26.self_attn.1.kv_a_proj_with_mqa": {
2387
+ "act_bits": 16,
2388
+ "bits": 8,
2389
+ "data_type": "int",
2390
+ "group_size": 128,
2391
+ "sym": true
2392
+ },
2393
+ "model.layers.26.self_attn.1.kv_b_proj": {
2394
+ "act_bits": 16,
2395
+ "bits": 8,
2396
+ "data_type": "int",
2397
+ "group_size": 128,
2398
+ "sym": true
2399
+ },
2400
+ "model.layers.26.self_attn.1.o_proj": {
2401
+ "act_bits": 16,
2402
+ "bits": 8,
2403
+ "data_type": "int",
2404
+ "group_size": 128,
2405
+ "sym": true
2406
+ },
2407
+ "model.layers.26.self_attn.1.q_a_proj": {
2408
+ "act_bits": 16,
2409
+ "bits": 8,
2410
+ "data_type": "int",
2411
+ "group_size": 128,
2412
+ "sym": true
2413
+ },
2414
+ "model.layers.26.self_attn.1.q_b_proj": {
2415
+ "act_bits": 16,
2416
+ "bits": 8,
2417
+ "data_type": "int",
2418
+ "group_size": 128,
2419
+ "sym": true
2420
+ },
2421
+ "model.layers.27.mlp.router.classifier": {
2422
+ "act_bits": 16,
2423
+ "bits": 8,
2424
+ "data_type": "int",
2425
+ "group_size": 128,
2426
+ "sym": true
2427
+ },
2428
+ "model.layers.27.mlps.0.down_proj": {
2429
+ "act_bits": 16,
2430
+ "bits": 8,
2431
+ "data_type": "int",
2432
+ "group_size": 128,
2433
+ "sym": true
2434
+ },
2435
+ "model.layers.27.mlps.0.gate_proj": {
2436
+ "act_bits": 16,
2437
+ "bits": 8,
2438
+ "data_type": "int",
2439
+ "group_size": 128,
2440
+ "sym": true
2441
+ },
2442
+ "model.layers.27.mlps.0.up_proj": {
2443
+ "act_bits": 16,
2444
+ "bits": 8,
2445
+ "data_type": "int",
2446
+ "group_size": 128,
2447
+ "sym": true
2448
+ },
2449
+ "model.layers.27.mlps.1.down_proj": {
2450
+ "act_bits": 16,
2451
+ "bits": 8,
2452
+ "data_type": "int",
2453
+ "group_size": 128,
2454
+ "sym": true
2455
+ },
2456
+ "model.layers.27.mlps.1.gate_proj": {
2457
+ "act_bits": 16,
2458
+ "bits": 8,
2459
+ "data_type": "int",
2460
+ "group_size": 128,
2461
+ "sym": true
2462
+ },
2463
+ "model.layers.27.mlps.1.up_proj": {
2464
+ "act_bits": 16,
2465
+ "bits": 8,
2466
+ "data_type": "int",
2467
+ "group_size": 128,
2468
+ "sym": true
2469
+ },
2470
+ "model.layers.27.self_attn.0.kv_a_proj_with_mqa": {
2471
+ "act_bits": 16,
2472
+ "bits": 8,
2473
+ "data_type": "int",
2474
+ "group_size": 128,
2475
+ "sym": true
2476
+ },
2477
+ "model.layers.27.self_attn.0.kv_b_proj": {
2478
+ "act_bits": 16,
2479
+ "bits": 8,
2480
+ "data_type": "int",
2481
+ "group_size": 128,
2482
+ "sym": true
2483
+ },
2484
+ "model.layers.27.self_attn.0.o_proj": {
2485
+ "act_bits": 16,
2486
+ "bits": 8,
2487
+ "data_type": "int",
2488
+ "group_size": 128,
2489
+ "sym": true
2490
+ },
2491
+ "model.layers.27.self_attn.0.q_a_proj": {
2492
+ "act_bits": 16,
2493
+ "bits": 8,
2494
+ "data_type": "int",
2495
+ "group_size": 128,
2496
+ "sym": true
2497
+ },
2498
+ "model.layers.27.self_attn.0.q_b_proj": {
2499
+ "act_bits": 16,
2500
+ "bits": 8,
2501
+ "data_type": "int",
2502
+ "group_size": 128,
2503
+ "sym": true
2504
+ },
2505
+ "model.layers.27.self_attn.1.kv_a_proj_with_mqa": {
2506
+ "act_bits": 16,
2507
+ "bits": 8,
2508
+ "data_type": "int",
2509
+ "group_size": 128,
2510
+ "sym": true
2511
+ },
2512
+ "model.layers.27.self_attn.1.kv_b_proj": {
2513
+ "act_bits": 16,
2514
+ "bits": 8,
2515
+ "data_type": "int",
2516
+ "group_size": 128,
2517
+ "sym": true
2518
+ },
2519
+ "model.layers.27.self_attn.1.o_proj": {
2520
+ "act_bits": 16,
2521
+ "bits": 8,
2522
+ "data_type": "int",
2523
+ "group_size": 128,
2524
+ "sym": true
2525
+ },
2526
+ "model.layers.27.self_attn.1.q_a_proj": {
2527
+ "act_bits": 16,
2528
+ "bits": 8,
2529
+ "data_type": "int",
2530
+ "group_size": 128,
2531
+ "sym": true
2532
+ },
2533
+ "model.layers.27.self_attn.1.q_b_proj": {
2534
+ "act_bits": 16,
2535
+ "bits": 8,
2536
+ "data_type": "int",
2537
+ "group_size": 128,
2538
+ "sym": true
2539
+ },
2540
+ "model.layers.3.mlp.router.classifier": {
2541
+ "act_bits": 16,
2542
+ "bits": 8,
2543
+ "data_type": "int",
2544
+ "group_size": 128,
2545
+ "sym": true
2546
+ },
2547
+ "model.layers.3.mlps.0.down_proj": {
2548
+ "act_bits": 16,
2549
+ "bits": 8,
2550
+ "data_type": "int",
2551
+ "group_size": 128,
2552
+ "sym": true
2553
+ },
2554
+ "model.layers.3.mlps.0.gate_proj": {
2555
+ "act_bits": 16,
2556
+ "bits": 8,
2557
+ "data_type": "int",
2558
+ "group_size": 128,
2559
+ "sym": true
2560
+ },
2561
+ "model.layers.3.mlps.0.up_proj": {
2562
+ "act_bits": 16,
2563
+ "bits": 8,
2564
+ "data_type": "int",
2565
+ "group_size": 128,
2566
+ "sym": true
2567
+ },
2568
+ "model.layers.3.mlps.1.down_proj": {
2569
+ "act_bits": 16,
2570
+ "bits": 8,
2571
+ "data_type": "int",
2572
+ "group_size": 128,
2573
+ "sym": true
2574
+ },
2575
+ "model.layers.3.mlps.1.gate_proj": {
2576
+ "act_bits": 16,
2577
+ "bits": 8,
2578
+ "data_type": "int",
2579
+ "group_size": 128,
2580
+ "sym": true
2581
+ },
2582
+ "model.layers.3.mlps.1.up_proj": {
2583
+ "act_bits": 16,
2584
+ "bits": 8,
2585
+ "data_type": "int",
2586
+ "group_size": 128,
2587
+ "sym": true
2588
+ },
2589
+ "model.layers.3.self_attn.0.kv_a_proj_with_mqa": {
2590
+ "act_bits": 16,
2591
+ "bits": 8,
2592
+ "data_type": "int",
2593
+ "group_size": 128,
2594
+ "sym": true
2595
+ },
2596
+ "model.layers.3.self_attn.0.kv_b_proj": {
2597
+ "act_bits": 16,
2598
+ "bits": 8,
2599
+ "data_type": "int",
2600
+ "group_size": 128,
2601
+ "sym": true
2602
+ },
2603
+ "model.layers.3.self_attn.0.o_proj": {
2604
+ "act_bits": 16,
2605
+ "bits": 8,
2606
+ "data_type": "int",
2607
+ "group_size": 128,
2608
+ "sym": true
2609
+ },
2610
+ "model.layers.3.self_attn.0.q_a_proj": {
2611
+ "act_bits": 16,
2612
+ "bits": 8,
2613
+ "data_type": "int",
2614
+ "group_size": 128,
2615
+ "sym": true
2616
+ },
2617
+ "model.layers.3.self_attn.0.q_b_proj": {
2618
+ "act_bits": 16,
2619
+ "bits": 8,
2620
+ "data_type": "int",
2621
+ "group_size": 128,
2622
+ "sym": true
2623
+ },
2624
+ "model.layers.3.self_attn.1.kv_a_proj_with_mqa": {
2625
+ "act_bits": 16,
2626
+ "bits": 8,
2627
+ "data_type": "int",
2628
+ "group_size": 128,
2629
+ "sym": true
2630
+ },
2631
+ "model.layers.3.self_attn.1.kv_b_proj": {
2632
+ "act_bits": 16,
2633
+ "bits": 8,
2634
+ "data_type": "int",
2635
+ "group_size": 128,
2636
+ "sym": true
2637
+ },
2638
+ "model.layers.3.self_attn.1.o_proj": {
2639
+ "act_bits": 16,
2640
+ "bits": 8,
2641
+ "data_type": "int",
2642
+ "group_size": 128,
2643
+ "sym": true
2644
+ },
2645
+ "model.layers.3.self_attn.1.q_a_proj": {
2646
+ "act_bits": 16,
2647
+ "bits": 8,
2648
+ "data_type": "int",
2649
+ "group_size": 128,
2650
+ "sym": true
2651
+ },
2652
+ "model.layers.3.self_attn.1.q_b_proj": {
2653
+ "act_bits": 16,
2654
+ "bits": 8,
2655
+ "data_type": "int",
2656
+ "group_size": 128,
2657
+ "sym": true
2658
+ },
2659
+ "model.layers.4.mlp.router.classifier": {
2660
+ "act_bits": 16,
2661
+ "bits": 8,
2662
+ "data_type": "int",
2663
+ "group_size": 128,
2664
+ "sym": true
2665
+ },
2666
+ "model.layers.4.mlps.0.down_proj": {
2667
+ "act_bits": 16,
2668
+ "bits": 8,
2669
+ "data_type": "int",
2670
+ "group_size": 128,
2671
+ "sym": true
2672
+ },
2673
+ "model.layers.4.mlps.0.gate_proj": {
2674
+ "act_bits": 16,
2675
+ "bits": 8,
2676
+ "data_type": "int",
2677
+ "group_size": 128,
2678
+ "sym": true
2679
+ },
2680
+ "model.layers.4.mlps.0.up_proj": {
2681
+ "act_bits": 16,
2682
+ "bits": 8,
2683
+ "data_type": "int",
2684
+ "group_size": 128,
2685
+ "sym": true
2686
+ },
2687
+ "model.layers.4.mlps.1.down_proj": {
2688
+ "act_bits": 16,
2689
+ "bits": 8,
2690
+ "data_type": "int",
2691
+ "group_size": 128,
2692
+ "sym": true
2693
+ },
2694
+ "model.layers.4.mlps.1.gate_proj": {
2695
+ "act_bits": 16,
2696
+ "bits": 8,
2697
+ "data_type": "int",
2698
+ "group_size": 128,
2699
+ "sym": true
2700
+ },
2701
+ "model.layers.4.mlps.1.up_proj": {
2702
+ "act_bits": 16,
2703
+ "bits": 8,
2704
+ "data_type": "int",
2705
+ "group_size": 128,
2706
+ "sym": true
2707
+ },
2708
+ "model.layers.4.self_attn.0.kv_a_proj_with_mqa": {
2709
+ "act_bits": 16,
2710
+ "bits": 8,
2711
+ "data_type": "int",
2712
+ "group_size": 128,
2713
+ "sym": true
2714
+ },
2715
+ "model.layers.4.self_attn.0.kv_b_proj": {
2716
+ "act_bits": 16,
2717
+ "bits": 8,
2718
+ "data_type": "int",
2719
+ "group_size": 128,
2720
+ "sym": true
2721
+ },
2722
+ "model.layers.4.self_attn.0.o_proj": {
2723
+ "act_bits": 16,
2724
+ "bits": 8,
2725
+ "data_type": "int",
2726
+ "group_size": 128,
2727
+ "sym": true
2728
+ },
2729
+ "model.layers.4.self_attn.0.q_a_proj": {
2730
+ "act_bits": 16,
2731
+ "bits": 8,
2732
+ "data_type": "int",
2733
+ "group_size": 128,
2734
+ "sym": true
2735
+ },
2736
+ "model.layers.4.self_attn.0.q_b_proj": {
2737
+ "act_bits": 16,
2738
+ "bits": 8,
2739
+ "data_type": "int",
2740
+ "group_size": 128,
2741
+ "sym": true
2742
+ },
2743
+ "model.layers.4.self_attn.1.kv_a_proj_with_mqa": {
2744
+ "act_bits": 16,
2745
+ "bits": 8,
2746
+ "data_type": "int",
2747
+ "group_size": 128,
2748
+ "sym": true
2749
+ },
2750
+ "model.layers.4.self_attn.1.kv_b_proj": {
2751
+ "act_bits": 16,
2752
+ "bits": 8,
2753
+ "data_type": "int",
2754
+ "group_size": 128,
2755
+ "sym": true
2756
+ },
2757
+ "model.layers.4.self_attn.1.o_proj": {
2758
+ "act_bits": 16,
2759
+ "bits": 8,
2760
+ "data_type": "int",
2761
+ "group_size": 128,
2762
+ "sym": true
2763
+ },
2764
+ "model.layers.4.self_attn.1.q_a_proj": {
2765
+ "act_bits": 16,
2766
+ "bits": 8,
2767
+ "data_type": "int",
2768
+ "group_size": 128,
2769
+ "sym": true
2770
+ },
2771
+ "model.layers.4.self_attn.1.q_b_proj": {
2772
+ "act_bits": 16,
2773
+ "bits": 8,
2774
+ "data_type": "int",
2775
+ "group_size": 128,
2776
+ "sym": true
2777
+ },
2778
+ "model.layers.5.mlp.router.classifier": {
2779
+ "act_bits": 16,
2780
+ "bits": 8,
2781
+ "data_type": "int",
2782
+ "group_size": 128,
2783
+ "sym": true
2784
+ },
2785
+ "model.layers.5.mlps.0.down_proj": {
2786
+ "act_bits": 16,
2787
+ "bits": 8,
2788
+ "data_type": "int",
2789
+ "group_size": 128,
2790
+ "sym": true
2791
+ },
2792
+ "model.layers.5.mlps.0.gate_proj": {
2793
+ "act_bits": 16,
2794
+ "bits": 8,
2795
+ "data_type": "int",
2796
+ "group_size": 128,
2797
+ "sym": true
2798
+ },
2799
+ "model.layers.5.mlps.0.up_proj": {
2800
+ "act_bits": 16,
2801
+ "bits": 8,
2802
+ "data_type": "int",
2803
+ "group_size": 128,
2804
+ "sym": true
2805
+ },
2806
+ "model.layers.5.mlps.1.down_proj": {
2807
+ "act_bits": 16,
2808
+ "bits": 8,
2809
+ "data_type": "int",
2810
+ "group_size": 128,
2811
+ "sym": true
2812
+ },
2813
+ "model.layers.5.mlps.1.gate_proj": {
2814
+ "act_bits": 16,
2815
+ "bits": 8,
2816
+ "data_type": "int",
2817
+ "group_size": 128,
2818
+ "sym": true
2819
+ },
2820
+ "model.layers.5.mlps.1.up_proj": {
2821
+ "act_bits": 16,
2822
+ "bits": 8,
2823
+ "data_type": "int",
2824
+ "group_size": 128,
2825
+ "sym": true
2826
+ },
2827
+ "model.layers.5.self_attn.0.kv_a_proj_with_mqa": {
2828
+ "act_bits": 16,
2829
+ "bits": 8,
2830
+ "data_type": "int",
2831
+ "group_size": 128,
2832
+ "sym": true
2833
+ },
2834
+ "model.layers.5.self_attn.0.kv_b_proj": {
2835
+ "act_bits": 16,
2836
+ "bits": 8,
2837
+ "data_type": "int",
2838
+ "group_size": 128,
2839
+ "sym": true
2840
+ },
2841
+ "model.layers.5.self_attn.0.o_proj": {
2842
+ "act_bits": 16,
2843
+ "bits": 8,
2844
+ "data_type": "int",
2845
+ "group_size": 128,
2846
+ "sym": true
2847
+ },
2848
+ "model.layers.5.self_attn.0.q_a_proj": {
2849
+ "act_bits": 16,
2850
+ "bits": 8,
2851
+ "data_type": "int",
2852
+ "group_size": 128,
2853
+ "sym": true
2854
+ },
2855
+ "model.layers.5.self_attn.0.q_b_proj": {
2856
+ "act_bits": 16,
2857
+ "bits": 8,
2858
+ "data_type": "int",
2859
+ "group_size": 128,
2860
+ "sym": true
2861
+ },
2862
+ "model.layers.5.self_attn.1.kv_a_proj_with_mqa": {
2863
+ "act_bits": 16,
2864
+ "bits": 8,
2865
+ "data_type": "int",
2866
+ "group_size": 128,
2867
+ "sym": true
2868
+ },
2869
+ "model.layers.5.self_attn.1.kv_b_proj": {
2870
+ "act_bits": 16,
2871
+ "bits": 8,
2872
+ "data_type": "int",
2873
+ "group_size": 128,
2874
+ "sym": true
2875
+ },
2876
+ "model.layers.5.self_attn.1.o_proj": {
2877
+ "act_bits": 16,
2878
+ "bits": 8,
2879
+ "data_type": "int",
2880
+ "group_size": 128,
2881
+ "sym": true
2882
+ },
2883
+ "model.layers.5.self_attn.1.q_a_proj": {
2884
+ "act_bits": 16,
2885
+ "bits": 8,
2886
+ "data_type": "int",
2887
+ "group_size": 128,
2888
+ "sym": true
2889
+ },
2890
+ "model.layers.5.self_attn.1.q_b_proj": {
2891
+ "act_bits": 16,
2892
+ "bits": 8,
2893
+ "data_type": "int",
2894
+ "group_size": 128,
2895
+ "sym": true
2896
+ },
2897
+ "model.layers.6.mlp.router.classifier": {
2898
+ "act_bits": 16,
2899
+ "bits": 8,
2900
+ "data_type": "int",
2901
+ "group_size": 128,
2902
+ "sym": true
2903
+ },
2904
+ "model.layers.6.mlps.0.down_proj": {
2905
+ "act_bits": 16,
2906
+ "bits": 8,
2907
+ "data_type": "int",
2908
+ "group_size": 128,
2909
+ "sym": true
2910
+ },
2911
+ "model.layers.6.mlps.0.gate_proj": {
2912
+ "act_bits": 16,
2913
+ "bits": 8,
2914
+ "data_type": "int",
2915
+ "group_size": 128,
2916
+ "sym": true
2917
+ },
2918
+ "model.layers.6.mlps.0.up_proj": {
2919
+ "act_bits": 16,
2920
+ "bits": 8,
2921
+ "data_type": "int",
2922
+ "group_size": 128,
2923
+ "sym": true
2924
+ },
2925
+ "model.layers.6.mlps.1.down_proj": {
2926
+ "act_bits": 16,
2927
+ "bits": 8,
2928
+ "data_type": "int",
2929
+ "group_size": 128,
2930
+ "sym": true
2931
+ },
2932
+ "model.layers.6.mlps.1.gate_proj": {
2933
+ "act_bits": 16,
2934
+ "bits": 8,
2935
+ "data_type": "int",
2936
+ "group_size": 128,
2937
+ "sym": true
2938
+ },
2939
+ "model.layers.6.mlps.1.up_proj": {
2940
+ "act_bits": 16,
2941
+ "bits": 8,
2942
+ "data_type": "int",
2943
+ "group_size": 128,
2944
+ "sym": true
2945
+ },
2946
+ "model.layers.6.self_attn.0.kv_a_proj_with_mqa": {
2947
+ "act_bits": 16,
2948
+ "bits": 8,
2949
+ "data_type": "int",
2950
+ "group_size": 128,
2951
+ "sym": true
2952
+ },
2953
+ "model.layers.6.self_attn.0.kv_b_proj": {
2954
+ "act_bits": 16,
2955
+ "bits": 8,
2956
+ "data_type": "int",
2957
+ "group_size": 128,
2958
+ "sym": true
2959
+ },
2960
+ "model.layers.6.self_attn.0.o_proj": {
2961
+ "act_bits": 16,
2962
+ "bits": 8,
2963
+ "data_type": "int",
2964
+ "group_size": 128,
2965
+ "sym": true
2966
+ },
2967
+ "model.layers.6.self_attn.0.q_a_proj": {
2968
+ "act_bits": 16,
2969
+ "bits": 8,
2970
+ "data_type": "int",
2971
+ "group_size": 128,
2972
+ "sym": true
2973
+ },
2974
+ "model.layers.6.self_attn.0.q_b_proj": {
2975
+ "act_bits": 16,
2976
+ "bits": 8,
2977
+ "data_type": "int",
2978
+ "group_size": 128,
2979
+ "sym": true
2980
+ },
2981
+ "model.layers.6.self_attn.1.kv_a_proj_with_mqa": {
2982
+ "act_bits": 16,
2983
+ "bits": 8,
2984
+ "data_type": "int",
2985
+ "group_size": 128,
2986
+ "sym": true
2987
+ },
2988
+ "model.layers.6.self_attn.1.kv_b_proj": {
2989
+ "act_bits": 16,
2990
+ "bits": 8,
2991
+ "data_type": "int",
2992
+ "group_size": 128,
2993
+ "sym": true
2994
+ },
2995
+ "model.layers.6.self_attn.1.o_proj": {
2996
+ "act_bits": 16,
2997
+ "bits": 8,
2998
+ "data_type": "int",
2999
+ "group_size": 128,
3000
+ "sym": true
3001
+ },
3002
+ "model.layers.6.self_attn.1.q_a_proj": {
3003
+ "act_bits": 16,
3004
+ "bits": 8,
3005
+ "data_type": "int",
3006
+ "group_size": 128,
3007
+ "sym": true
3008
+ },
3009
+ "model.layers.6.self_attn.1.q_b_proj": {
3010
+ "act_bits": 16,
3011
+ "bits": 8,
3012
+ "data_type": "int",
3013
+ "group_size": 128,
3014
+ "sym": true
3015
+ },
3016
+ "model.layers.7.mlp.router.classifier": {
3017
+ "act_bits": 16,
3018
+ "bits": 8,
3019
+ "data_type": "int",
3020
+ "group_size": 128,
3021
+ "sym": true
3022
+ },
3023
+ "model.layers.7.mlps.0.down_proj": {
3024
+ "act_bits": 16,
3025
+ "bits": 8,
3026
+ "data_type": "int",
3027
+ "group_size": 128,
3028
+ "sym": true
3029
+ },
3030
+ "model.layers.7.mlps.0.gate_proj": {
3031
+ "act_bits": 16,
3032
+ "bits": 8,
3033
+ "data_type": "int",
3034
+ "group_size": 128,
3035
+ "sym": true
3036
+ },
3037
+ "model.layers.7.mlps.0.up_proj": {
3038
+ "act_bits": 16,
3039
+ "bits": 8,
3040
+ "data_type": "int",
3041
+ "group_size": 128,
3042
+ "sym": true
3043
+ },
3044
+ "model.layers.7.mlps.1.down_proj": {
3045
+ "act_bits": 16,
3046
+ "bits": 8,
3047
+ "data_type": "int",
3048
+ "group_size": 128,
3049
+ "sym": true
3050
+ },
3051
+ "model.layers.7.mlps.1.gate_proj": {
3052
+ "act_bits": 16,
3053
+ "bits": 8,
3054
+ "data_type": "int",
3055
+ "group_size": 128,
3056
+ "sym": true
3057
+ },
3058
+ "model.layers.7.mlps.1.up_proj": {
3059
+ "act_bits": 16,
3060
+ "bits": 8,
3061
+ "data_type": "int",
3062
+ "group_size": 128,
3063
+ "sym": true
3064
+ },
3065
+ "model.layers.7.self_attn.0.kv_a_proj_with_mqa": {
3066
+ "act_bits": 16,
3067
+ "bits": 8,
3068
+ "data_type": "int",
3069
+ "group_size": 128,
3070
+ "sym": true
3071
+ },
3072
+ "model.layers.7.self_attn.0.kv_b_proj": {
3073
+ "act_bits": 16,
3074
+ "bits": 8,
3075
+ "data_type": "int",
3076
+ "group_size": 128,
3077
+ "sym": true
3078
+ },
3079
+ "model.layers.7.self_attn.0.o_proj": {
3080
+ "act_bits": 16,
3081
+ "bits": 8,
3082
+ "data_type": "int",
3083
+ "group_size": 128,
3084
+ "sym": true
3085
+ },
3086
+ "model.layers.7.self_attn.0.q_a_proj": {
3087
+ "act_bits": 16,
3088
+ "bits": 8,
3089
+ "data_type": "int",
3090
+ "group_size": 128,
3091
+ "sym": true
3092
+ },
3093
+ "model.layers.7.self_attn.0.q_b_proj": {
3094
+ "act_bits": 16,
3095
+ "bits": 8,
3096
+ "data_type": "int",
3097
+ "group_size": 128,
3098
+ "sym": true
3099
+ },
3100
+ "model.layers.7.self_attn.1.kv_a_proj_with_mqa": {
3101
+ "act_bits": 16,
3102
+ "bits": 8,
3103
+ "data_type": "int",
3104
+ "group_size": 128,
3105
+ "sym": true
3106
+ },
3107
+ "model.layers.7.self_attn.1.kv_b_proj": {
3108
+ "act_bits": 16,
3109
+ "bits": 8,
3110
+ "data_type": "int",
3111
+ "group_size": 128,
3112
+ "sym": true
3113
+ },
3114
+ "model.layers.7.self_attn.1.o_proj": {
3115
+ "act_bits": 16,
3116
+ "bits": 8,
3117
+ "data_type": "int",
3118
+ "group_size": 128,
3119
+ "sym": true
3120
+ },
3121
+ "model.layers.7.self_attn.1.q_a_proj": {
3122
+ "act_bits": 16,
3123
+ "bits": 8,
3124
+ "data_type": "int",
3125
+ "group_size": 128,
3126
+ "sym": true
3127
+ },
3128
+ "model.layers.7.self_attn.1.q_b_proj": {
3129
+ "act_bits": 16,
3130
+ "bits": 8,
3131
+ "data_type": "int",
3132
+ "group_size": 128,
3133
+ "sym": true
3134
+ },
3135
+ "model.layers.8.mlp.router.classifier": {
3136
+ "act_bits": 16,
3137
+ "bits": 8,
3138
+ "data_type": "int",
3139
+ "group_size": 128,
3140
+ "sym": true
3141
+ },
3142
+ "model.layers.8.mlps.0.down_proj": {
3143
+ "act_bits": 16,
3144
+ "bits": 8,
3145
+ "data_type": "int",
3146
+ "group_size": 128,
3147
+ "sym": true
3148
+ },
3149
+ "model.layers.8.mlps.0.gate_proj": {
3150
+ "act_bits": 16,
3151
+ "bits": 8,
3152
+ "data_type": "int",
3153
+ "group_size": 128,
3154
+ "sym": true
3155
+ },
3156
+ "model.layers.8.mlps.0.up_proj": {
3157
+ "act_bits": 16,
3158
+ "bits": 8,
3159
+ "data_type": "int",
3160
+ "group_size": 128,
3161
+ "sym": true
3162
+ },
3163
+ "model.layers.8.mlps.1.down_proj": {
3164
+ "act_bits": 16,
3165
+ "bits": 8,
3166
+ "data_type": "int",
3167
+ "group_size": 128,
3168
+ "sym": true
3169
+ },
3170
+ "model.layers.8.mlps.1.gate_proj": {
3171
+ "act_bits": 16,
3172
+ "bits": 8,
3173
+ "data_type": "int",
3174
+ "group_size": 128,
3175
+ "sym": true
3176
+ },
3177
+ "model.layers.8.mlps.1.up_proj": {
3178
+ "act_bits": 16,
3179
+ "bits": 8,
3180
+ "data_type": "int",
3181
+ "group_size": 128,
3182
+ "sym": true
3183
+ },
3184
+ "model.layers.8.self_attn.0.kv_a_proj_with_mqa": {
3185
+ "act_bits": 16,
3186
+ "bits": 8,
3187
+ "data_type": "int",
3188
+ "group_size": 128,
3189
+ "sym": true
3190
+ },
3191
+ "model.layers.8.self_attn.0.kv_b_proj": {
3192
+ "act_bits": 16,
3193
+ "bits": 8,
3194
+ "data_type": "int",
3195
+ "group_size": 128,
3196
+ "sym": true
3197
+ },
3198
+ "model.layers.8.self_attn.0.o_proj": {
3199
+ "act_bits": 16,
3200
+ "bits": 8,
3201
+ "data_type": "int",
3202
+ "group_size": 128,
3203
+ "sym": true
3204
+ },
3205
+ "model.layers.8.self_attn.0.q_a_proj": {
3206
+ "act_bits": 16,
3207
+ "bits": 8,
3208
+ "data_type": "int",
3209
+ "group_size": 128,
3210
+ "sym": true
3211
+ },
3212
+ "model.layers.8.self_attn.0.q_b_proj": {
3213
+ "act_bits": 16,
3214
+ "bits": 8,
3215
+ "data_type": "int",
3216
+ "group_size": 128,
3217
+ "sym": true
3218
+ },
3219
+ "model.layers.8.self_attn.1.kv_a_proj_with_mqa": {
3220
+ "act_bits": 16,
3221
+ "bits": 8,
3222
+ "data_type": "int",
3223
+ "group_size": 128,
3224
+ "sym": true
3225
+ },
3226
+ "model.layers.8.self_attn.1.kv_b_proj": {
3227
+ "act_bits": 16,
3228
+ "bits": 8,
3229
+ "data_type": "int",
3230
+ "group_size": 128,
3231
+ "sym": true
3232
+ },
3233
+ "model.layers.8.self_attn.1.o_proj": {
3234
+ "act_bits": 16,
3235
+ "bits": 8,
3236
+ "data_type": "int",
3237
+ "group_size": 128,
3238
+ "sym": true
3239
+ },
3240
+ "model.layers.8.self_attn.1.q_a_proj": {
3241
+ "act_bits": 16,
3242
+ "bits": 8,
3243
+ "data_type": "int",
3244
+ "group_size": 128,
3245
+ "sym": true
3246
+ },
3247
+ "model.layers.8.self_attn.1.q_b_proj": {
3248
+ "act_bits": 16,
3249
+ "bits": 8,
3250
+ "data_type": "int",
3251
+ "group_size": 128,
3252
+ "sym": true
3253
+ },
3254
+ "model.layers.9.mlp.router.classifier": {
3255
+ "act_bits": 16,
3256
+ "bits": 8,
3257
+ "data_type": "int",
3258
+ "group_size": 128,
3259
+ "sym": true
3260
+ },
3261
+ "model.layers.9.mlps.0.down_proj": {
3262
+ "act_bits": 16,
3263
+ "bits": 8,
3264
+ "data_type": "int",
3265
+ "group_size": 128,
3266
+ "sym": true
3267
+ },
3268
+ "model.layers.9.mlps.0.gate_proj": {
3269
+ "act_bits": 16,
3270
+ "bits": 8,
3271
+ "data_type": "int",
3272
+ "group_size": 128,
3273
+ "sym": true
3274
+ },
3275
+ "model.layers.9.mlps.0.up_proj": {
3276
+ "act_bits": 16,
3277
+ "bits": 8,
3278
+ "data_type": "int",
3279
+ "group_size": 128,
3280
+ "sym": true
3281
+ },
3282
+ "model.layers.9.mlps.1.down_proj": {
3283
+ "act_bits": 16,
3284
+ "bits": 8,
3285
+ "data_type": "int",
3286
+ "group_size": 128,
3287
+ "sym": true
3288
+ },
3289
+ "model.layers.9.mlps.1.gate_proj": {
3290
+ "act_bits": 16,
3291
+ "bits": 8,
3292
+ "data_type": "int",
3293
+ "group_size": 128,
3294
+ "sym": true
3295
+ },
3296
+ "model.layers.9.mlps.1.up_proj": {
3297
+ "act_bits": 16,
3298
+ "bits": 8,
3299
+ "data_type": "int",
3300
+ "group_size": 128,
3301
+ "sym": true
3302
+ },
3303
+ "model.layers.9.self_attn.0.kv_a_proj_with_mqa": {
3304
+ "act_bits": 16,
3305
+ "bits": 8,
3306
+ "data_type": "int",
3307
+ "group_size": 128,
3308
+ "sym": true
3309
+ },
3310
+ "model.layers.9.self_attn.0.kv_b_proj": {
3311
+ "act_bits": 16,
3312
+ "bits": 8,
3313
+ "data_type": "int",
3314
+ "group_size": 128,
3315
+ "sym": true
3316
+ },
3317
+ "model.layers.9.self_attn.0.o_proj": {
3318
+ "act_bits": 16,
3319
+ "bits": 8,
3320
+ "data_type": "int",
3321
+ "group_size": 128,
3322
+ "sym": true
3323
+ },
3324
+ "model.layers.9.self_attn.0.q_a_proj": {
3325
+ "act_bits": 16,
3326
+ "bits": 8,
3327
+ "data_type": "int",
3328
+ "group_size": 128,
3329
+ "sym": true
3330
+ },
3331
+ "model.layers.9.self_attn.0.q_b_proj": {
3332
+ "act_bits": 16,
3333
+ "bits": 8,
3334
+ "data_type": "int",
3335
+ "group_size": 128,
3336
+ "sym": true
3337
+ },
3338
+ "model.layers.9.self_attn.1.kv_a_proj_with_mqa": {
3339
+ "act_bits": 16,
3340
+ "bits": 8,
3341
+ "data_type": "int",
3342
+ "group_size": 128,
3343
+ "sym": true
3344
+ },
3345
+ "model.layers.9.self_attn.1.kv_b_proj": {
3346
+ "act_bits": 16,
3347
+ "bits": 8,
3348
+ "data_type": "int",
3349
+ "group_size": 128,
3350
+ "sym": true
3351
+ },
3352
+ "model.layers.9.self_attn.1.o_proj": {
3353
+ "act_bits": 16,
3354
+ "bits": 8,
3355
+ "data_type": "int",
3356
+ "group_size": 128,
3357
+ "sym": true
3358
+ },
3359
+ "model.layers.9.self_attn.1.q_a_proj": {
3360
+ "act_bits": 16,
3361
+ "bits": 8,
3362
+ "data_type": "int",
3363
+ "group_size": 128,
3364
+ "sym": true
3365
+ },
3366
+ "model.layers.9.self_attn.1.q_b_proj": {
3367
+ "act_bits": 16,
3368
+ "bits": 8,
3369
+ "data_type": "int",
3370
+ "group_size": 128,
3371
+ "sym": true
3372
+ }
3373
+ },
3374
+ "group_size": 128,
3375
+ "iters": 0,
3376
+ "packing_format": "auto_round:auto_gptq",
3377
+ "quant_method": "auto-round",
3378
+ "sym": true
3379
+ },
3380
+ "rms_norm_eps": 1e-05,
3381
+ "rope_scaling": {
3382
+ "beta_fast": 32,
3383
+ "beta_slow": 1,
3384
+ "factor": 120,
3385
+ "mscale": 1,
3386
+ "mscale_all_dim": 1,
3387
+ "original_max_position_embeddings": 8192,
3388
+ "rope_type": "yarn"
3389
+ },
3390
+ "rope_theta": 1000000.0,
3391
+ "routed_scaling_factor": 6.0,
3392
+ "router_bias": false,
3393
+ "tie_word_embeddings": false,
3394
+ "torch_dtype": "bfloat16",
3395
+ "transformers_version": "4.53.3",
3396
+ "use_cache": true,
3397
+ "v_head_dim": 128,
3398
+ "vocab_size": 131072,
3399
+ "zero_expert_num": 256,
3400
+ "zero_expert_type": "identity"
3401
+ }
configuration_longcat.py ADDED
@@ -0,0 +1,210 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2025 bzantium and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+
5
+ """Longcat model configuration"""
6
+
7
+ from transformers.configuration_utils import PretrainedConfig
8
+ from transformers.modeling_rope_utils import rope_config_validation
9
+
10
+
11
+ LONGCAT_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
12
+
13
+
14
+ class LongcatConfig(PretrainedConfig):
15
+ r"""
16
+ This is the configuration class to store the configuration of a [`LongcatModel`]. It is used to instantiate an Longcat
17
+ model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
18
+ defaults will yield a similar configuration to that of the Longcat.
19
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
20
+ documentation from [`PretrainedConfig`] for more information.
21
+
22
+
23
+ Args:
24
+ vocab_size (`int`, *optional*, defaults to 131072):
25
+ Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
26
+ `inputs_ids` passed when calling [`LongcatModel`]
27
+ hidden_size (`int`, *optional*, defaults to 7168):
28
+ Dimension of the hidden representations.
29
+ ffn_hidden_size (`int`, *optional*, defaults to 18432):
30
+ Dimension of the MLP representations.
31
+ expert_ffn_hidden_size (`int`, *optional*, defaults to 2048):
32
+ Dimension of the MoE representations.
33
+ num_layers (`int`, *optional*, defaults to 61):
34
+ Number of hidden layers in the Transformer decoder.
35
+ num_attention_heads (`int`, *optional*, defaults to 128):
36
+ Number of attention heads for each attention layer in the Transformer decoder.
37
+ num_key_value_heads (`int`, *optional*, defaults to 128):
38
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
39
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
40
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
41
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
42
+ by meanpooling all the original heads within that group. For more details checkout [this
43
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
44
+ `num_attention_heads`.
45
+ n_routed_experts (`int`, *optional*, defaults to 256):
46
+ Number of routed experts.
47
+ routed_scaling_factor (`float`, *optional*, defaults to 2.5):
48
+ Scaling factor or routed experts.
49
+ kv_lora_rank (`int`, *optional*, defaults to 512):
50
+ Rank of the LoRA matrices for key and value projections.
51
+ q_lora_rank (`int`, *optional*, defaults to 1536):
52
+ Rank of the LoRA matrices for query projections.
53
+ qk_rope_head_dim (`int`, *optional*, defaults to 64):
54
+ Dimension of the query/key heads that use rotary position embeddings.
55
+ v_head_dim (`int`, *optional*, defaults to 128):
56
+ Dimension of the value heads.
57
+ qk_nope_head_dim (`int`, *optional*, defaults to 128):
58
+ Dimension of the query/key heads that don't use rotary position embeddings.
59
+ norm_topk_prob (`bool`, *optional*, defaults to `True`):
60
+ Whether to normalize the weights of the routed experts.
61
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
62
+ The non-linear activation function (function or string) in the decoder.
63
+ max_position_embeddings (`int`, *optional*, defaults to 4096):
64
+ The maximum sequence length that this model might ever be used with.
65
+ rms_norm_eps (`float`, *optional*, defaults to 1e-06):
66
+ The epsilon used by the rms normalization layers.
67
+ use_cache (`bool`, *optional*, defaults to `True`):
68
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
69
+ relevant if `config.is_decoder=True`.
70
+ pad_token_id (`int`, *optional*):
71
+ Padding token id.
72
+ bos_token_id (`int`, *optional*, defaults to 0):
73
+ Beginning of stream token id.
74
+ eos_token_id (`int`, *optional*, defaults to 1):
75
+ End of stream token id.
76
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
77
+ Whether to tie weight embeddings
78
+ rope_theta (`float`, *optional*, defaults to 10000.0):
79
+ The base period of the RoPE embeddings.
80
+ attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
81
+ Whether to use a bias in the query, key, value and output projection layers during self-attention.
82
+ attention_dropout (`float`, *optional*, defaults to 0.0):
83
+ The dropout ratio for the attention probabilities.
84
+
85
+ ```python
86
+ >>> from transformers import LongcatModel, LongcatConfig
87
+
88
+ >>> # Initializing a Longcat style configuration
89
+ >>> configuration = LongcatConfig()
90
+
91
+ >>> # Accessing the model configuration
92
+ >>> configuration = model.config
93
+ ```"""
94
+
95
+ model_type = "longcat"
96
+ keys_to_ignore_at_inference = ["past_key_values"]
97
+ base_model_tp_plan = { # TODO: only replicate attention layers when > first_k_dense_replace
98
+ "layers.*.self_attn.k_proj": "colwise",
99
+ "layers.*.self_attn.v_proj": "colwise",
100
+ "layers.*.self_attn.o_proj": "rowwise",
101
+ "layers.*.mlp.experts.*.gate_proj": "local_colwise",
102
+ "layers.*.mlp.experts.*.up_proj": "local_colwise",
103
+ "layers.*.mlp.experts.*.down_proj": "local_rowwise",
104
+ "layers.*.mlps.*.gate_proj": "local_colwise",
105
+ "layers.*.mlps.*.up_proj": "local_colwise",
106
+ "layers.*.mlps.*.down_proj": "local_rowwise",
107
+ }
108
+ base_model_pp_plan = {
109
+ "embed_tokens": (["input_ids"], ["inputs_embeds"]),
110
+ "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
111
+ "norm": (["hidden_states"], ["hidden_states"]),
112
+ }
113
+
114
+ def __init__(
115
+ self,
116
+ vocab_size=131072,
117
+ hidden_size=7168,
118
+ ffn_hidden_size=18432,
119
+ expert_ffn_hidden_size=2048,
120
+ num_layers=61,
121
+ num_attention_heads=128,
122
+ num_key_value_heads=None,
123
+ n_routed_experts=256,
124
+ routed_scaling_factor=1,
125
+ kv_lora_rank=512,
126
+ q_lora_rank=1536,
127
+ qk_rope_head_dim=64,
128
+ v_head_dim=128,
129
+ head_dim=128,
130
+ qk_nope_head_dim=128,
131
+ mla_scale_q_lora=True,
132
+ mla_scale_kv_lora=True,
133
+ moe_topk=8,
134
+ norm_topk_prob=False,
135
+ hidden_act="silu",
136
+ max_position_embeddings=4096,
137
+ rms_norm_eps=1e-6,
138
+ use_cache=True,
139
+ pad_token_id=None,
140
+ bos_token_id=0,
141
+ eos_token_id=1,
142
+ tie_word_embeddings=False,
143
+ rope_theta=10000.0,
144
+ attention_bias=False,
145
+ attention_dropout=0.0,
146
+ attention_method='GQA',
147
+ initializer_range=0.006,
148
+ router_bias=False,
149
+ zero_expert_num=None,
150
+ zero_expert_type=None,
151
+ **kwargs,
152
+ ):
153
+ self.vocab_size = vocab_size
154
+ self.max_position_embeddings = max_position_embeddings
155
+ self.hidden_size = hidden_size
156
+ self.ffn_hidden_size = ffn_hidden_size
157
+ self.expert_ffn_hidden_size = expert_ffn_hidden_size
158
+ self.num_layers = num_layers
159
+ self.num_attention_heads = num_attention_heads
160
+ self.n_routed_experts = n_routed_experts
161
+ self.routed_scaling_factor = routed_scaling_factor
162
+ self.kv_lora_rank = kv_lora_rank
163
+ self.q_lora_rank = q_lora_rank
164
+ self.qk_rope_head_dim = qk_rope_head_dim
165
+ self.v_head_dim = v_head_dim
166
+ self.qk_nope_head_dim = qk_nope_head_dim
167
+ self.qk_head_dim = qk_nope_head_dim + qk_rope_head_dim
168
+ self.moe_topk = moe_topk
169
+ self.norm_topk_prob = norm_topk_prob
170
+ self.mla_scale_q_lora = mla_scale_q_lora
171
+ self.mla_scale_kv_lora = mla_scale_kv_lora
172
+ self.attention_method = attention_method
173
+ self.initializer_range = initializer_range
174
+ self.router_bias = router_bias
175
+ self.zero_expert_num = zero_expert_num
176
+ self.zero_expert_type = zero_expert_type
177
+
178
+ if self.attention_method == "GQA":
179
+ self.head_dim = head_dim
180
+ elif self.attention_method == "MLA":
181
+ self.head_dim = qk_rope_head_dim
182
+ else:
183
+ ValueError("attention_method should be one of [\"GQA\", \"MLA\"]")
184
+
185
+
186
+ # for backward compatibility
187
+ if num_key_value_heads is None:
188
+ num_key_value_heads = num_attention_heads
189
+
190
+ self.num_key_value_heads = num_key_value_heads
191
+ self.hidden_act = hidden_act
192
+ self.rms_norm_eps = rms_norm_eps
193
+ self.use_cache = use_cache
194
+ self.rope_theta = rope_theta
195
+ self.attention_bias = attention_bias
196
+ self.attention_dropout = attention_dropout
197
+ # Validate the correctness of rotary position embeddings parameters
198
+ # BC: if there is a 'type' field, copy it it to 'rope_type'.
199
+ rope_config_validation(self)
200
+
201
+ super().__init__(
202
+ pad_token_id=pad_token_id,
203
+ bos_token_id=bos_token_id,
204
+ eos_token_id=eos_token_id,
205
+ tie_word_embeddings=tie_word_embeddings,
206
+ **kwargs,
207
+ )
208
+
209
+
210
+ __all__ = ["LongcatConfig"]
generation_config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 1,
4
+ "eos_token_id": 2,
5
+ "pad_token_id": 0,
6
+ "transformers_version": "4.53.3"
7
+ }
model-00001-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4684e5c7f85fe61b479fd209f5fd4ae93f951885cab85a1ee757e3bd2a6a54cf
3
+ size 4997079704
model-00002-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f078d5e7bcd9ef6733a20182225aaf7b9e27cc59c3cff1d3b4fc6a89c76be035
3
+ size 4994713520
model-00003-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:43d02bed6dafc4e1334cd2c60cf7f569717f7cd95b0062508c5ff0bc351257cc
3
+ size 4994837976
model-00004-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:adfd78db66a8242bf5a32d5d0f6b028a24c5a25408bf6f61f4ed72c91596d2c6
3
+ size 4994713520
model-00005-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d49176fc6b3671651f67f94850370834813a212ed700a91e74c3de37537b6dd
3
+ size 4994837976
model-00006-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e5aeb3f3337e94e1bf0f5e7509fcc842bcff2c7bee1bba7e9242e41962832045
3
+ size 4994713520
model-00007-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:36466d48d5edc97cf3a76142cdc5d01030769842eef2d7ba64917698deec6cf8
3
+ size 4994838296
model-00008-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f6c8273ae424e9d396f42494b0b0fe89b6c897448004802a55c1baacf63968b5
3
+ size 4994713208
model-00009-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d981cf843ec2d7fae26ad08d860800ca3c22812113183eb9a221444a51fb9278
3
+ size 4994838616
model-00010-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:af06441fdb705d31473610f55bb58f4dc03607d500e178d2ed3cb28c3d16aaa2
3
+ size 4994712880
model-00011-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0c6c717aef80208f00a380123622a02ace04a7a7da82b0b2c97c0c0bf59665c9
3
+ size 4984081616
model-00012-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57cbb24c1ccf7679cd9f9f45618752b954eb41d03e1937afe2a5912fddd6169f
3
+ size 4998932440
model-00013-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:65b44a539f072c0769160dc30ca28891a2932dae8c4d9c3785de3375675d2a1e
3
+ size 4994713520
model-00014-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e01b088d9549be6c444e0f692ecec71dd4c89b8ba9184baef36b0dcafaf42bbd
3
+ size 4994838320
model-00015-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5445750ab08b7e01249c11fe90ce879653807eba5c6fc3cacfcff76ce66925c9
3
+ size 4994713520
model-00016-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4b02b4c1ebd5b7427934142ec19988a8f71587dd7a0d7d3f3a3a5c25c9c0e28b
3
+ size 4994838048
model-00017-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:21b59c6e6857b09fdfbc89556455ecf2a819f79b41d3f64fccc949cf95d0a553
3
+ size 4994713520
model-00018-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:21b21568b9b92054203df88b5bae4da191ca5c3fccf267f4b5e0f6c497152add
3
+ size 4994837984
model-00019-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7be2eb1f8016989abe50005d2766a9346e65e8c222cb0667c275ba049d3c6458
3
+ size 4994713520
model-00020-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:66eca59142fa91e054960f63f53f1766ee52f261c58dacdfc5bb01943dfc7e04
3
+ size 4994837984
model-00021-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bbf49263a59ae7d8c92b9090c19c923f1a9ee3227df3e0326c5e31c257b19dc1
3
+ size 4994713520
model-00022-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c28a516a40d9590937fc3089ade4ceeee8a7537197b24f1680bb0a80fbaad01
3
+ size 4994838880
model-00023-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91800f05f5f9c97845c6149d97fa599573733f4ece45cb0734a1c3515c1713e5
3
+ size 4994715512
model-00024-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a3ed027e8b89b7211f234bb47e922dc73defb3ae6d8667e6aa763d818600c653
3
+ size 4994840656
model-00025-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0413b6c03dc22963b543b55eaa24f9fc757e24596f8b9e4d1105301dca33b01
3
+ size 4994715184
model-00026-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:53e201e958133f95c45041443ce76a1eca26ba4a0a58aef47416181b72f68e67
3
+ size 4938320624
model-00027-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ee33fb43099de0f438eccb1d24d7d941c1ef37c89cb804110d22c47c5a8347f4
3
+ size 4998934672
model-00028-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:932cb33be8adbfff2f82c9c6370698582d8b37afa0f581c7d7bba6ffe8bf42f1
3
+ size 4994715808
model-00029-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cda3b5139bedf8f1f0be069b682b4a339bc280544335d9a29afa1de1e6284bed
3
+ size 4994840368
model-00030-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ab2b79ca710de42adba808e544f53f65efff7964751cb7e8916af04a28f6807b
3
+ size 4994715808
model-00031-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4ed47b1db1ae5ddbd50a54e5ca93a2f6e15ce07dc4efcf3b50d84f9254d0c73
3
+ size 4994840096
model-00032-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b2922f5de9a704bb93a083fb6564070fd2ef450b632bbb6add6865818d5ac320
3
+ size 4994715808
model-00033-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:40e97cf44ba6d1027dfb3213402380a6c5d4ef451934233a11161d33d7976c69
3
+ size 4994840032
model-00034-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18d06c755a278d7e5815f00ffde4090ddffb463a13a4e2258574688f416ff140
3
+ size 4994715808
model-00035-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ea7ba424a4846c69bfd08bcec57f9061370ce6977dd6d44d02096ccf48d956d
3
+ size 4994840032
model-00036-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:986155abc7dcc1b4506c8bb863790c42574f1b4138d7ad856e10d241bb43a579
3
+ size 4994715808
model-00037-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cdc005d94768c7f8ca78ee526967aa804f86b49e84ff91153c8e9517f7d34453
3
+ size 4994840328
model-00038-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fad3fd5661aef8d82b0b97c26b935816caf6cf094166969fa06ab067a66627e6
3
+ size 4994715512
model-00039-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:993f0113cf108afd6c56ee3ddefd906155f4338c2b2091ccb346158e71025842
3
+ size 4994840656
model-00040-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:87d5848deb628f404dac7ec74a45963d956037a2ad0473af4686e6a0dd45ed68
3
+ size 4994715184
model-00041-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a9bdb5d4e9a80cba58b4fbd232505b8cf659864c6f7406e756b8e5e6eb047987
3
+ size 4938320624
model-00042-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3987c6fa9c4d98a804fd2a2afbc77c372e29cb705ae0913ffdfa5b2b4bd6074
3
+ size 4998934672
model-00043-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1557d62e0222ebb4a1d679ee3cdcbb0de6492320ac7a3fb03d6ff982df59487
3
+ size 4994715808
model-00044-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a4bdf3f368b90b16998fa8a3e0ace222411bf42b91b9f0afc9351bdbc66aac70
3
+ size 4994840368
model-00045-of-00061.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a59ba42049797c6c0e8ef1cc43922623c9cea98ad2b994755ca391864b7039a8
3
+ size 4994715808