Hugging Face's logo

INC4AI
/
GLM-5.1-int4-mixed-AutoRound

Safetensors
glm_moe_dsa
4-bit precision
auto-round
Model card Files
xet
 Community
1
INC4AI commited on
Commit
83e108a
·
verified ·
1 Parent(s): 5b0b8be

Upload model

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 +2 -0
  2. README.md +0 -0
  3. chat_template.jinja +117 -0
  4. config.json +2839 -0
  5. generation_config.json +13 -0
  6. model-00001-of-00075.safetensors +3 -0
  7. model-00002-of-00075.safetensors +3 -0
  8. model-00003-of-00075.safetensors +3 -0
  9. model-00004-of-00075.safetensors +3 -0
  10. model-00005-of-00075.safetensors +3 -0
  11. model-00006-of-00075.safetensors +3 -0
  12. model-00007-of-00075.safetensors +3 -0
  13. model-00008-of-00075.safetensors +3 -0
  14. model-00009-of-00075.safetensors +3 -0
  15. model-00010-of-00075.safetensors +3 -0
  16. model-00011-of-00075.safetensors +3 -0
  17. model-00012-of-00075.safetensors +3 -0
  18. model-00013-of-00075.safetensors +3 -0
  19. model-00014-of-00075.safetensors +3 -0
  20. model-00015-of-00075.safetensors +3 -0
  21. model-00016-of-00075.safetensors +3 -0
  22. model-00017-of-00075.safetensors +3 -0
  23. model-00018-of-00075.safetensors +3 -0
  24. model-00019-of-00075.safetensors +3 -0
  25. model-00020-of-00075.safetensors +3 -0
  26. model-00021-of-00075.safetensors +3 -0
  27. model-00022-of-00075.safetensors +3 -0
  28. model-00023-of-00075.safetensors +3 -0
  29. model-00024-of-00075.safetensors +3 -0
  30. model-00025-of-00075.safetensors +3 -0
  31. model-00026-of-00075.safetensors +3 -0
  32. model-00027-of-00075.safetensors +3 -0
  33. model-00028-of-00075.safetensors +3 -0
  34. model-00029-of-00075.safetensors +3 -0
  35. model-00030-of-00075.safetensors +3 -0
  36. model-00031-of-00075.safetensors +3 -0
  37. model-00032-of-00075.safetensors +3 -0
  38. model-00033-of-00075.safetensors +3 -0
  39. model-00034-of-00075.safetensors +3 -0
  40. model-00035-of-00075.safetensors +3 -0
  41. model-00036-of-00075.safetensors +3 -0
  42. model-00037-of-00075.safetensors +3 -0
  43. model-00038-of-00075.safetensors +3 -0
  44. model-00039-of-00075.safetensors +3 -0
  45. model-00040-of-00075.safetensors +3 -0
  46. model-00041-of-00075.safetensors +3 -0
  47. model-00042-of-00075.safetensors +3 -0
  48. model-00043-of-00075.safetensors +3 -0
  49. model-00044-of-00075.safetensors +3 -0
  50. model-00045-of-00075.safetensors +3 -0
.gitattributes CHANGED
@@ -33,3 +33,5 @@ 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
37
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
README.md ADDED
File without changes
chat_template.jinja ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [gMASK]<sop>
2
+ {%- if tools -%}
3
+ {%- macro tool_to_json(tool) -%}
4
+ {%- set ns_tool = namespace(first=true) -%}
5
+ {{ '{' -}}
6
+ {%- for k, v in tool.items() -%}
7
+ {%- if k != 'defer_loading' and k != 'strict' -%}
8
+ {%- if not ns_tool.first -%}{{- ', ' -}}{%- endif -%}
9
+ {%- set ns_tool.first = false -%}
10
+ "{{ k }}": {{ v | tojson(ensure_ascii=False) }}
11
+ {%- endif -%}
12
+ {%- endfor -%}
13
+ {{- '}' -}}
14
+ {%- endmacro -%}
15
+ <|system|>
16
+ # Tools
17
+
18
+ You may call one or more functions to assist with the user query.
19
+
20
+ You are provided with function signatures within <tools></tools> XML tags:
21
+ <tools>
22
+ {% for tool in tools %}
23
+ {%- if 'function' in tool -%}
24
+ {%- set tool = tool['function'] -%}
25
+ {%- endif -%}
26
+ {% if tool.defer_loading is not defined or not tool.defer_loading %}
27
+ {{ tool_to_json(tool) }}
28
+ {% endif %}
29
+ {% endfor %}
30
+ </tools>
31
+
32
+ For each function call, output the function name and arguments within the following XML format:
33
+ <tool_call>{function-name}<arg_key>{arg-key-1}</arg_key><arg_value>{arg-value-1}</arg_value><arg_key>{arg-key-2}</arg_key><arg_value>{arg-value-2}</arg_value>...</tool_call>{%- endif -%}
34
+ {%- macro visible_text(content) -%}
35
+ {%- if content is string -%}
36
+ {{- content }}
37
+ {%- elif content is iterable and content is not mapping -%}
38
+ {%- for item in content -%}
39
+ {%- if item is mapping and item.type == 'text' -%}
40
+ {{- item.text }}
41
+ {%- elif item is string -%}
42
+ {{- item }}
43
+ {%- endif -%}
44
+ {%- endfor -%}
45
+ {%- else -%}
46
+ {{- content }}
47
+ {%- endif -%}
48
+ {%- endmacro -%}
49
+ {%- set ns = namespace(last_user_index=-1, thinking_indices='') -%}
50
+ {%- for m in messages %}
51
+ {%- if m.role == 'user' %}
52
+ {%- set ns.last_user_index = loop.index0 -%}
53
+ {%- elif m.role == 'assistant' %}
54
+ {%- if m.reasoning_content is string %}
55
+ {%- set ns.thinking_indices = ns.thinking_indices ~ ',' ~ ns.last_user_index ~ ',' -%}
56
+ {%- endif %}
57
+ {%- endif %}
58
+ {%- endfor %}
59
+ {%- set ns.has_thinking = false -%}
60
+ {%- for m in messages -%}
61
+ {%- if m.role == 'user' -%}<|user|>{{ visible_text(m.content) }}{% set ns.has_thinking = (',' ~ loop.index0 ~ ',') in ns.thinking_indices -%}
62
+ {%- elif m.role == 'assistant' -%}
63
+ <|assistant|>
64
+ {%- set content = visible_text(m.content) %}
65
+ {%- if m.reasoning_content is string %}
66
+ {%- set reasoning_content = m.reasoning_content %}
67
+ {%- elif '</think>' in content %}
68
+ {%- set reasoning_content = content.split('</think>')[0].split('<think>')[-1] %}
69
+ {%- set content = content.split('</think>')[-1] %}
70
+ {%- elif loop.index0 > ns.last_user_index and not (enable_thinking is defined and not enable_thinking) %}
71
+ {%- set reasoning_content = '' %}
72
+ {%- elif loop.index0 < ns.last_user_index and ns.has_thinking %}
73
+ {%- set reasoning_content = '' %}
74
+ {%- endif %}
75
+ {%- if ((clear_thinking is defined and not clear_thinking) or loop.index0 > ns.last_user_index) and reasoning_content is defined -%}
76
+ {{ '<think>' + reasoning_content + '</think>'}}
77
+ {%- else -%}
78
+ {{ '</think>' }}
79
+ {%- endif -%}
80
+ {%- if content.strip() -%}
81
+ {{ content.strip() }}
82
+ {%- endif -%}
83
+ {% if m.tool_calls %}
84
+ {% for tc in m.tool_calls %}
85
+ {%- if tc.function %}
86
+ {%- set tc = tc.function %}
87
+ {%- endif %}
88
+ {{- '<tool_call>' + tc.name -}}
89
+ {% set _args = tc.arguments %}{% for k, v in _args.items() %}<arg_key>{{ k }}</arg_key><arg_value>{{ v | tojson(ensure_ascii=False) if v is not string else v }}</arg_value>{% endfor %}</tool_call>{% endfor %}
90
+ {% endif %}
91
+ {%- elif m.role == 'tool' -%}
92
+ {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
93
+ {{- '<|observation|>' -}}
94
+ {%- endif %}
95
+ {%- if m.content is string -%}
96
+ {{- '<tool_response>' + m.content + '</tool_response>' -}}
97
+ {%- else -%}
98
+ {{- '<tool_response><tools>\n' -}}
99
+ {% for tr in m.content %}
100
+ {%- for tool in tools -%}
101
+ {%- if 'function' in tool -%}
102
+ {%- set tool = tool['function'] -%}
103
+ {%- endif -%}
104
+ {%- if tool.name == tr.name -%}
105
+ {{- tool_to_json(tool) + '\n' -}}
106
+ {%- endif -%}
107
+ {%- endfor -%}
108
+ {%- endfor -%}
109
+ {{- '</tools></tool_response>' -}}
110
+ {% endif -%}
111
+ {%- elif m.role == 'system' -%}
112
+ <|system|>{{ visible_text(m.content) }}
113
+ {%- endif -%}
114
+ {%- endfor -%}
115
+ {%- if add_generation_prompt -%}
116
+ <|assistant|>{{- '</think>' if (enable_thinking is defined and not enable_thinking) else '<think>' -}}
117
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,2839 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "GlmMoeDsaForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 0,
8
+ "dtype": "bfloat16",
9
+ "eos_token_id": [
10
+ 154820,
11
+ 154827,
12
+ 154829
13
+ ],
14
+ "ep_size": 1,
15
+ "first_k_dense_replace": 3,
16
+ "hidden_act": "silu",
17
+ "hidden_size": 6144,
18
+ "index_head_dim": 128,
19
+ "index_n_heads": 32,
20
+ "index_topk": 2048,
21
+ "indexer_rope_interleave": true,
22
+ "initializer_range": 0.02,
23
+ "intermediate_size": 12288,
24
+ "kv_lora_rank": 512,
25
+ "max_position_embeddings": 202752,
26
+ "mlp_layer_types": [
27
+ "dense",
28
+ "dense",
29
+ "dense",
30
+ "sparse",
31
+ "sparse",
32
+ "sparse",
33
+ "sparse",
34
+ "sparse",
35
+ "sparse",
36
+ "sparse",
37
+ "sparse",
38
+ "sparse",
39
+ "sparse",
40
+ "sparse",
41
+ "sparse",
42
+ "sparse",
43
+ "sparse",
44
+ "sparse",
45
+ "sparse",
46
+ "sparse",
47
+ "sparse",
48
+ "sparse",
49
+ "sparse",
50
+ "sparse",
51
+ "sparse",
52
+ "sparse",
53
+ "sparse",
54
+ "sparse",
55
+ "sparse",
56
+ "sparse",
57
+ "sparse",
58
+ "sparse",
59
+ "sparse",
60
+ "sparse",
61
+ "sparse",
62
+ "sparse",
63
+ "sparse",
64
+ "sparse",
65
+ "sparse",
66
+ "sparse",
67
+ "sparse",
68
+ "sparse",
69
+ "sparse",
70
+ "sparse",
71
+ "sparse",
72
+ "sparse",
73
+ "sparse",
74
+ "sparse",
75
+ "sparse",
76
+ "sparse",
77
+ "sparse",
78
+ "sparse",
79
+ "sparse",
80
+ "sparse",
81
+ "sparse",
82
+ "sparse",
83
+ "sparse",
84
+ "sparse",
85
+ "sparse",
86
+ "sparse",
87
+ "sparse",
88
+ "sparse",
89
+ "sparse",
90
+ "sparse",
91
+ "sparse",
92
+ "sparse",
93
+ "sparse",
94
+ "sparse",
95
+ "sparse",
96
+ "sparse",
97
+ "sparse",
98
+ "sparse",
99
+ "sparse",
100
+ "sparse",
101
+ "sparse",
102
+ "sparse",
103
+ "sparse",
104
+ "sparse"
105
+ ],
106
+ "model_type": "glm_moe_dsa",
107
+ "moe_intermediate_size": 2048,
108
+ "moe_layer_freq": 1,
109
+ "n_group": 1,
110
+ "n_routed_experts": 256,
111
+ "n_shared_experts": 1,
112
+ "norm_topk_prob": true,
113
+ "num_attention_heads": 64,
114
+ "num_experts_per_tok": 8,
115
+ "num_hidden_layers": 78,
116
+ "num_key_value_heads": 64,
117
+ "num_nextn_predict_layers": 1,
118
+ "pad_token_id": 154820,
119
+ "pretraining_tp": 1,
120
+ "q_lora_rank": 2048,
121
+ "qk_head_dim": 256,
122
+ "qk_nope_head_dim": 192,
123
+ "qk_rope_head_dim": 64,
124
+ "quantization_config": {
125
+ "autoround_version": "0.13.0",
126
+ "bits": 4,
127
+ "data_type": "int",
128
+ "extra_config": {
129
+ ".*layers\\.0\\.mlp.*": {
130
+ "bits": 16,
131
+ "data_type": "float"
132
+ },
133
+ ".*layers\\.1\\.mlp.*": {
134
+ "bits": 16,
135
+ "data_type": "float"
136
+ },
137
+ ".*layers\\.2\\.mlp.*": {
138
+ "bits": 16,
139
+ "data_type": "float"
140
+ },
141
+ ".*weights_proj.*": {
142
+ "bits": 16,
143
+ "data_type": "float"
144
+ },
145
+ "model.layers.0.mlp.down_proj": {
146
+ "bits": 16,
147
+ "data_type": "float"
148
+ },
149
+ "model.layers.0.mlp.gate_proj": {
150
+ "bits": 16,
151
+ "data_type": "float"
152
+ },
153
+ "model.layers.0.mlp.up_proj": {
154
+ "bits": 16,
155
+ "data_type": "float"
156
+ },
157
+ "model.layers.0.self_attn.indexer.weights_proj": {
158
+ "bits": 16,
159
+ "data_type": "float"
160
+ },
161
+ "model.layers.0.self_attn.indexer.wk": {
162
+ "bits": 8
163
+ },
164
+ "model.layers.0.self_attn.indexer.wq_b": {
165
+ "bits": 8
166
+ },
167
+ "model.layers.0.self_attn.kv_a_proj_with_mqa": {
168
+ "bits": 8
169
+ },
170
+ "model.layers.0.self_attn.kv_b_proj": {
171
+ "bits": 8
172
+ },
173
+ "model.layers.0.self_attn.o_proj": {
174
+ "bits": 8
175
+ },
176
+ "model.layers.0.self_attn.q_a_proj": {
177
+ "bits": 8
178
+ },
179
+ "model.layers.0.self_attn.q_b_proj": {
180
+ "bits": 8
181
+ },
182
+ "model.layers.1.mlp.down_proj": {
183
+ "bits": 16,
184
+ "data_type": "float"
185
+ },
186
+ "model.layers.1.mlp.gate_proj": {
187
+ "bits": 16,
188
+ "data_type": "float"
189
+ },
190
+ "model.layers.1.mlp.up_proj": {
191
+ "bits": 16,
192
+ "data_type": "float"
193
+ },
194
+ "model.layers.1.self_attn.indexer.weights_proj": {
195
+ "bits": 16,
196
+ "data_type": "float"
197
+ },
198
+ "model.layers.1.self_attn.indexer.wk": {
199
+ "bits": 8
200
+ },
201
+ "model.layers.1.self_attn.indexer.wq_b": {
202
+ "bits": 8
203
+ },
204
+ "model.layers.1.self_attn.kv_a_proj_with_mqa": {
205
+ "bits": 8
206
+ },
207
+ "model.layers.1.self_attn.kv_b_proj": {
208
+ "bits": 8
209
+ },
210
+ "model.layers.1.self_attn.o_proj": {
211
+ "bits": 8
212
+ },
213
+ "model.layers.1.self_attn.q_a_proj": {
214
+ "bits": 8
215
+ },
216
+ "model.layers.1.self_attn.q_b_proj": {
217
+ "bits": 8
218
+ },
219
+ "model.layers.10.mlp.shared_experts.down_proj": {
220
+ "bits": 8
221
+ },
222
+ "model.layers.10.mlp.shared_experts.gate_proj": {
223
+ "bits": 8
224
+ },
225
+ "model.layers.10.mlp.shared_experts.up_proj": {
226
+ "bits": 8
227
+ },
228
+ "model.layers.10.self_attn.indexer.weights_proj": {
229
+ "bits": 16,
230
+ "data_type": "float"
231
+ },
232
+ "model.layers.10.self_attn.indexer.wk": {
233
+ "bits": 8
234
+ },
235
+ "model.layers.10.self_attn.indexer.wq_b": {
236
+ "bits": 8
237
+ },
238
+ "model.layers.10.self_attn.kv_a_proj_with_mqa": {
239
+ "bits": 8
240
+ },
241
+ "model.layers.10.self_attn.kv_b_proj": {
242
+ "bits": 8
243
+ },
244
+ "model.layers.10.self_attn.o_proj": {
245
+ "bits": 8
246
+ },
247
+ "model.layers.10.self_attn.q_a_proj": {
248
+ "bits": 8
249
+ },
250
+ "model.layers.10.self_attn.q_b_proj": {
251
+ "bits": 8
252
+ },
253
+ "model.layers.11.mlp.shared_experts.down_proj": {
254
+ "bits": 8
255
+ },
256
+ "model.layers.11.mlp.shared_experts.gate_proj": {
257
+ "bits": 8
258
+ },
259
+ "model.layers.11.mlp.shared_experts.up_proj": {
260
+ "bits": 8
261
+ },
262
+ "model.layers.11.self_attn.indexer.weights_proj": {
263
+ "bits": 16,
264
+ "data_type": "float"
265
+ },
266
+ "model.layers.11.self_attn.indexer.wk": {
267
+ "bits": 8
268
+ },
269
+ "model.layers.11.self_attn.indexer.wq_b": {
270
+ "bits": 8
271
+ },
272
+ "model.layers.11.self_attn.kv_a_proj_with_mqa": {
273
+ "bits": 8
274
+ },
275
+ "model.layers.11.self_attn.kv_b_proj": {
276
+ "bits": 8
277
+ },
278
+ "model.layers.11.self_attn.o_proj": {
279
+ "bits": 8
280
+ },
281
+ "model.layers.11.self_attn.q_a_proj": {
282
+ "bits": 8
283
+ },
284
+ "model.layers.11.self_attn.q_b_proj": {
285
+ "bits": 8
286
+ },
287
+ "model.layers.12.mlp.shared_experts.down_proj": {
288
+ "bits": 8
289
+ },
290
+ "model.layers.12.mlp.shared_experts.gate_proj": {
291
+ "bits": 8
292
+ },
293
+ "model.layers.12.mlp.shared_experts.up_proj": {
294
+ "bits": 8
295
+ },
296
+ "model.layers.12.self_attn.indexer.weights_proj": {
297
+ "bits": 16,
298
+ "data_type": "float"
299
+ },
300
+ "model.layers.12.self_attn.indexer.wk": {
301
+ "bits": 8
302
+ },
303
+ "model.layers.12.self_attn.indexer.wq_b": {
304
+ "bits": 8
305
+ },
306
+ "model.layers.12.self_attn.kv_a_proj_with_mqa": {
307
+ "bits": 8
308
+ },
309
+ "model.layers.12.self_attn.kv_b_proj": {
310
+ "bits": 8
311
+ },
312
+ "model.layers.12.self_attn.o_proj": {
313
+ "bits": 8
314
+ },
315
+ "model.layers.12.self_attn.q_a_proj": {
316
+ "bits": 8
317
+ },
318
+ "model.layers.12.self_attn.q_b_proj": {
319
+ "bits": 8
320
+ },
321
+ "model.layers.13.mlp.shared_experts.down_proj": {
322
+ "bits": 8
323
+ },
324
+ "model.layers.13.mlp.shared_experts.gate_proj": {
325
+ "bits": 8
326
+ },
327
+ "model.layers.13.mlp.shared_experts.up_proj": {
328
+ "bits": 8
329
+ },
330
+ "model.layers.13.self_attn.indexer.weights_proj": {
331
+ "bits": 16,
332
+ "data_type": "float"
333
+ },
334
+ "model.layers.13.self_attn.indexer.wk": {
335
+ "bits": 8
336
+ },
337
+ "model.layers.13.self_attn.indexer.wq_b": {
338
+ "bits": 8
339
+ },
340
+ "model.layers.13.self_attn.kv_a_proj_with_mqa": {
341
+ "bits": 8
342
+ },
343
+ "model.layers.13.self_attn.kv_b_proj": {
344
+ "bits": 8
345
+ },
346
+ "model.layers.13.self_attn.o_proj": {
347
+ "bits": 8
348
+ },
349
+ "model.layers.13.self_attn.q_a_proj": {
350
+ "bits": 8
351
+ },
352
+ "model.layers.13.self_attn.q_b_proj": {
353
+ "bits": 8
354
+ },
355
+ "model.layers.14.mlp.shared_experts.down_proj": {
356
+ "bits": 8
357
+ },
358
+ "model.layers.14.mlp.shared_experts.gate_proj": {
359
+ "bits": 8
360
+ },
361
+ "model.layers.14.mlp.shared_experts.up_proj": {
362
+ "bits": 8
363
+ },
364
+ "model.layers.14.self_attn.indexer.weights_proj": {
365
+ "bits": 16,
366
+ "data_type": "float"
367
+ },
368
+ "model.layers.14.self_attn.indexer.wk": {
369
+ "bits": 8
370
+ },
371
+ "model.layers.14.self_attn.indexer.wq_b": {
372
+ "bits": 8
373
+ },
374
+ "model.layers.14.self_attn.kv_a_proj_with_mqa": {
375
+ "bits": 8
376
+ },
377
+ "model.layers.14.self_attn.kv_b_proj": {
378
+ "bits": 8
379
+ },
380
+ "model.layers.14.self_attn.o_proj": {
381
+ "bits": 8
382
+ },
383
+ "model.layers.14.self_attn.q_a_proj": {
384
+ "bits": 8
385
+ },
386
+ "model.layers.14.self_attn.q_b_proj": {
387
+ "bits": 8
388
+ },
389
+ "model.layers.15.mlp.shared_experts.down_proj": {
390
+ "bits": 8
391
+ },
392
+ "model.layers.15.mlp.shared_experts.gate_proj": {
393
+ "bits": 8
394
+ },
395
+ "model.layers.15.mlp.shared_experts.up_proj": {
396
+ "bits": 8
397
+ },
398
+ "model.layers.15.self_attn.indexer.weights_proj": {
399
+ "bits": 16,
400
+ "data_type": "float"
401
+ },
402
+ "model.layers.15.self_attn.indexer.wk": {
403
+ "bits": 8
404
+ },
405
+ "model.layers.15.self_attn.indexer.wq_b": {
406
+ "bits": 8
407
+ },
408
+ "model.layers.15.self_attn.kv_a_proj_with_mqa": {
409
+ "bits": 8
410
+ },
411
+ "model.layers.15.self_attn.kv_b_proj": {
412
+ "bits": 8
413
+ },
414
+ "model.layers.15.self_attn.o_proj": {
415
+ "bits": 8
416
+ },
417
+ "model.layers.15.self_attn.q_a_proj": {
418
+ "bits": 8
419
+ },
420
+ "model.layers.15.self_attn.q_b_proj": {
421
+ "bits": 8
422
+ },
423
+ "model.layers.16.mlp.shared_experts.down_proj": {
424
+ "bits": 8
425
+ },
426
+ "model.layers.16.mlp.shared_experts.gate_proj": {
427
+ "bits": 8
428
+ },
429
+ "model.layers.16.mlp.shared_experts.up_proj": {
430
+ "bits": 8
431
+ },
432
+ "model.layers.16.self_attn.indexer.weights_proj": {
433
+ "bits": 16,
434
+ "data_type": "float"
435
+ },
436
+ "model.layers.16.self_attn.indexer.wk": {
437
+ "bits": 8
438
+ },
439
+ "model.layers.16.self_attn.indexer.wq_b": {
440
+ "bits": 8
441
+ },
442
+ "model.layers.16.self_attn.kv_a_proj_with_mqa": {
443
+ "bits": 8
444
+ },
445
+ "model.layers.16.self_attn.kv_b_proj": {
446
+ "bits": 8
447
+ },
448
+ "model.layers.16.self_attn.o_proj": {
449
+ "bits": 8
450
+ },
451
+ "model.layers.16.self_attn.q_a_proj": {
452
+ "bits": 8
453
+ },
454
+ "model.layers.16.self_attn.q_b_proj": {
455
+ "bits": 8
456
+ },
457
+ "model.layers.17.mlp.shared_experts.down_proj": {
458
+ "bits": 8
459
+ },
460
+ "model.layers.17.mlp.shared_experts.gate_proj": {
461
+ "bits": 8
462
+ },
463
+ "model.layers.17.mlp.shared_experts.up_proj": {
464
+ "bits": 8
465
+ },
466
+ "model.layers.17.self_attn.indexer.weights_proj": {
467
+ "bits": 16,
468
+ "data_type": "float"
469
+ },
470
+ "model.layers.17.self_attn.indexer.wk": {
471
+ "bits": 8
472
+ },
473
+ "model.layers.17.self_attn.indexer.wq_b": {
474
+ "bits": 8
475
+ },
476
+ "model.layers.17.self_attn.kv_a_proj_with_mqa": {
477
+ "bits": 8
478
+ },
479
+ "model.layers.17.self_attn.kv_b_proj": {
480
+ "bits": 8
481
+ },
482
+ "model.layers.17.self_attn.o_proj": {
483
+ "bits": 8
484
+ },
485
+ "model.layers.17.self_attn.q_a_proj": {
486
+ "bits": 8
487
+ },
488
+ "model.layers.17.self_attn.q_b_proj": {
489
+ "bits": 8
490
+ },
491
+ "model.layers.18.mlp.shared_experts.down_proj": {
492
+ "bits": 8
493
+ },
494
+ "model.layers.18.mlp.shared_experts.gate_proj": {
495
+ "bits": 8
496
+ },
497
+ "model.layers.18.mlp.shared_experts.up_proj": {
498
+ "bits": 8
499
+ },
500
+ "model.layers.18.self_attn.indexer.weights_proj": {
501
+ "bits": 16,
502
+ "data_type": "float"
503
+ },
504
+ "model.layers.18.self_attn.indexer.wk": {
505
+ "bits": 8
506
+ },
507
+ "model.layers.18.self_attn.indexer.wq_b": {
508
+ "bits": 8
509
+ },
510
+ "model.layers.18.self_attn.kv_a_proj_with_mqa": {
511
+ "bits": 8
512
+ },
513
+ "model.layers.18.self_attn.kv_b_proj": {
514
+ "bits": 8
515
+ },
516
+ "model.layers.18.self_attn.o_proj": {
517
+ "bits": 8
518
+ },
519
+ "model.layers.18.self_attn.q_a_proj": {
520
+ "bits": 8
521
+ },
522
+ "model.layers.18.self_attn.q_b_proj": {
523
+ "bits": 8
524
+ },
525
+ "model.layers.19.mlp.shared_experts.down_proj": {
526
+ "bits": 8
527
+ },
528
+ "model.layers.19.mlp.shared_experts.gate_proj": {
529
+ "bits": 8
530
+ },
531
+ "model.layers.19.mlp.shared_experts.up_proj": {
532
+ "bits": 8
533
+ },
534
+ "model.layers.19.self_attn.indexer.weights_proj": {
535
+ "bits": 16,
536
+ "data_type": "float"
537
+ },
538
+ "model.layers.19.self_attn.indexer.wk": {
539
+ "bits": 8
540
+ },
541
+ "model.layers.19.self_attn.indexer.wq_b": {
542
+ "bits": 8
543
+ },
544
+ "model.layers.19.self_attn.kv_a_proj_with_mqa": {
545
+ "bits": 8
546
+ },
547
+ "model.layers.19.self_attn.kv_b_proj": {
548
+ "bits": 8
549
+ },
550
+ "model.layers.19.self_attn.o_proj": {
551
+ "bits": 8
552
+ },
553
+ "model.layers.19.self_attn.q_a_proj": {
554
+ "bits": 8
555
+ },
556
+ "model.layers.19.self_attn.q_b_proj": {
557
+ "bits": 8
558
+ },
559
+ "model.layers.2.mlp.down_proj": {
560
+ "bits": 16,
561
+ "data_type": "float"
562
+ },
563
+ "model.layers.2.mlp.gate_proj": {
564
+ "bits": 16,
565
+ "data_type": "float"
566
+ },
567
+ "model.layers.2.mlp.up_proj": {
568
+ "bits": 16,
569
+ "data_type": "float"
570
+ },
571
+ "model.layers.2.self_attn.indexer.weights_proj": {
572
+ "bits": 16,
573
+ "data_type": "float"
574
+ },
575
+ "model.layers.2.self_attn.indexer.wk": {
576
+ "bits": 8
577
+ },
578
+ "model.layers.2.self_attn.indexer.wq_b": {
579
+ "bits": 8
580
+ },
581
+ "model.layers.2.self_attn.kv_a_proj_with_mqa": {
582
+ "bits": 8
583
+ },
584
+ "model.layers.2.self_attn.kv_b_proj": {
585
+ "bits": 8
586
+ },
587
+ "model.layers.2.self_attn.o_proj": {
588
+ "bits": 8
589
+ },
590
+ "model.layers.2.self_attn.q_a_proj": {
591
+ "bits": 8
592
+ },
593
+ "model.layers.2.self_attn.q_b_proj": {
594
+ "bits": 8
595
+ },
596
+ "model.layers.20.mlp.shared_experts.down_proj": {
597
+ "bits": 8
598
+ },
599
+ "model.layers.20.mlp.shared_experts.gate_proj": {
600
+ "bits": 8
601
+ },
602
+ "model.layers.20.mlp.shared_experts.up_proj": {
603
+ "bits": 8
604
+ },
605
+ "model.layers.20.self_attn.indexer.weights_proj": {
606
+ "bits": 16,
607
+ "data_type": "float"
608
+ },
609
+ "model.layers.20.self_attn.indexer.wk": {
610
+ "bits": 8
611
+ },
612
+ "model.layers.20.self_attn.indexer.wq_b": {
613
+ "bits": 8
614
+ },
615
+ "model.layers.20.self_attn.kv_a_proj_with_mqa": {
616
+ "bits": 8
617
+ },
618
+ "model.layers.20.self_attn.kv_b_proj": {
619
+ "bits": 8
620
+ },
621
+ "model.layers.20.self_attn.o_proj": {
622
+ "bits": 8
623
+ },
624
+ "model.layers.20.self_attn.q_a_proj": {
625
+ "bits": 8
626
+ },
627
+ "model.layers.20.self_attn.q_b_proj": {
628
+ "bits": 8
629
+ },
630
+ "model.layers.21.mlp.shared_experts.down_proj": {
631
+ "bits": 8
632
+ },
633
+ "model.layers.21.mlp.shared_experts.gate_proj": {
634
+ "bits": 8
635
+ },
636
+ "model.layers.21.mlp.shared_experts.up_proj": {
637
+ "bits": 8
638
+ },
639
+ "model.layers.21.self_attn.indexer.weights_proj": {
640
+ "bits": 16,
641
+ "data_type": "float"
642
+ },
643
+ "model.layers.21.self_attn.indexer.wk": {
644
+ "bits": 8
645
+ },
646
+ "model.layers.21.self_attn.indexer.wq_b": {
647
+ "bits": 8
648
+ },
649
+ "model.layers.21.self_attn.kv_a_proj_with_mqa": {
650
+ "bits": 8
651
+ },
652
+ "model.layers.21.self_attn.kv_b_proj": {
653
+ "bits": 8
654
+ },
655
+ "model.layers.21.self_attn.o_proj": {
656
+ "bits": 8
657
+ },
658
+ "model.layers.21.self_attn.q_a_proj": {
659
+ "bits": 8
660
+ },
661
+ "model.layers.21.self_attn.q_b_proj": {
662
+ "bits": 8
663
+ },
664
+ "model.layers.22.mlp.shared_experts.down_proj": {
665
+ "bits": 8
666
+ },
667
+ "model.layers.22.mlp.shared_experts.gate_proj": {
668
+ "bits": 8
669
+ },
670
+ "model.layers.22.mlp.shared_experts.up_proj": {
671
+ "bits": 8
672
+ },
673
+ "model.layers.22.self_attn.indexer.weights_proj": {
674
+ "bits": 16,
675
+ "data_type": "float"
676
+ },
677
+ "model.layers.22.self_attn.indexer.wk": {
678
+ "bits": 8
679
+ },
680
+ "model.layers.22.self_attn.indexer.wq_b": {
681
+ "bits": 8
682
+ },
683
+ "model.layers.22.self_attn.kv_a_proj_with_mqa": {
684
+ "bits": 8
685
+ },
686
+ "model.layers.22.self_attn.kv_b_proj": {
687
+ "bits": 8
688
+ },
689
+ "model.layers.22.self_attn.o_proj": {
690
+ "bits": 8
691
+ },
692
+ "model.layers.22.self_attn.q_a_proj": {
693
+ "bits": 8
694
+ },
695
+ "model.layers.22.self_attn.q_b_proj": {
696
+ "bits": 8
697
+ },
698
+ "model.layers.23.mlp.shared_experts.down_proj": {
699
+ "bits": 8
700
+ },
701
+ "model.layers.23.mlp.shared_experts.gate_proj": {
702
+ "bits": 8
703
+ },
704
+ "model.layers.23.mlp.shared_experts.up_proj": {
705
+ "bits": 8
706
+ },
707
+ "model.layers.23.self_attn.indexer.weights_proj": {
708
+ "bits": 16,
709
+ "data_type": "float"
710
+ },
711
+ "model.layers.23.self_attn.indexer.wk": {
712
+ "bits": 8
713
+ },
714
+ "model.layers.23.self_attn.indexer.wq_b": {
715
+ "bits": 8
716
+ },
717
+ "model.layers.23.self_attn.kv_a_proj_with_mqa": {
718
+ "bits": 8
719
+ },
720
+ "model.layers.23.self_attn.kv_b_proj": {
721
+ "bits": 8
722
+ },
723
+ "model.layers.23.self_attn.o_proj": {
724
+ "bits": 8
725
+ },
726
+ "model.layers.23.self_attn.q_a_proj": {
727
+ "bits": 8
728
+ },
729
+ "model.layers.23.self_attn.q_b_proj": {
730
+ "bits": 8
731
+ },
732
+ "model.layers.24.mlp.shared_experts.down_proj": {
733
+ "bits": 8
734
+ },
735
+ "model.layers.24.mlp.shared_experts.gate_proj": {
736
+ "bits": 8
737
+ },
738
+ "model.layers.24.mlp.shared_experts.up_proj": {
739
+ "bits": 8
740
+ },
741
+ "model.layers.24.self_attn.indexer.weights_proj": {
742
+ "bits": 16,
743
+ "data_type": "float"
744
+ },
745
+ "model.layers.24.self_attn.indexer.wk": {
746
+ "bits": 8
747
+ },
748
+ "model.layers.24.self_attn.indexer.wq_b": {
749
+ "bits": 8
750
+ },
751
+ "model.layers.24.self_attn.kv_a_proj_with_mqa": {
752
+ "bits": 8
753
+ },
754
+ "model.layers.24.self_attn.kv_b_proj": {
755
+ "bits": 8
756
+ },
757
+ "model.layers.24.self_attn.o_proj": {
758
+ "bits": 8
759
+ },
760
+ "model.layers.24.self_attn.q_a_proj": {
761
+ "bits": 8
762
+ },
763
+ "model.layers.24.self_attn.q_b_proj": {
764
+ "bits": 8
765
+ },
766
+ "model.layers.25.mlp.shared_experts.down_proj": {
767
+ "bits": 8
768
+ },
769
+ "model.layers.25.mlp.shared_experts.gate_proj": {
770
+ "bits": 8
771
+ },
772
+ "model.layers.25.mlp.shared_experts.up_proj": {
773
+ "bits": 8
774
+ },
775
+ "model.layers.25.self_attn.indexer.weights_proj": {
776
+ "bits": 16,
777
+ "data_type": "float"
778
+ },
779
+ "model.layers.25.self_attn.indexer.wk": {
780
+ "bits": 8
781
+ },
782
+ "model.layers.25.self_attn.indexer.wq_b": {
783
+ "bits": 8
784
+ },
785
+ "model.layers.25.self_attn.kv_a_proj_with_mqa": {
786
+ "bits": 8
787
+ },
788
+ "model.layers.25.self_attn.kv_b_proj": {
789
+ "bits": 8
790
+ },
791
+ "model.layers.25.self_attn.o_proj": {
792
+ "bits": 8
793
+ },
794
+ "model.layers.25.self_attn.q_a_proj": {
795
+ "bits": 8
796
+ },
797
+ "model.layers.25.self_attn.q_b_proj": {
798
+ "bits": 8
799
+ },
800
+ "model.layers.26.mlp.shared_experts.down_proj": {
801
+ "bits": 8
802
+ },
803
+ "model.layers.26.mlp.shared_experts.gate_proj": {
804
+ "bits": 8
805
+ },
806
+ "model.layers.26.mlp.shared_experts.up_proj": {
807
+ "bits": 8
808
+ },
809
+ "model.layers.26.self_attn.indexer.weights_proj": {
810
+ "bits": 16,
811
+ "data_type": "float"
812
+ },
813
+ "model.layers.26.self_attn.indexer.wk": {
814
+ "bits": 8
815
+ },
816
+ "model.layers.26.self_attn.indexer.wq_b": {
817
+ "bits": 8
818
+ },
819
+ "model.layers.26.self_attn.kv_a_proj_with_mqa": {
820
+ "bits": 8
821
+ },
822
+ "model.layers.26.self_attn.kv_b_proj": {
823
+ "bits": 8
824
+ },
825
+ "model.layers.26.self_attn.o_proj": {
826
+ "bits": 8
827
+ },
828
+ "model.layers.26.self_attn.q_a_proj": {
829
+ "bits": 8
830
+ },
831
+ "model.layers.26.self_attn.q_b_proj": {
832
+ "bits": 8
833
+ },
834
+ "model.layers.27.mlp.shared_experts.down_proj": {
835
+ "bits": 8
836
+ },
837
+ "model.layers.27.mlp.shared_experts.gate_proj": {
838
+ "bits": 8
839
+ },
840
+ "model.layers.27.mlp.shared_experts.up_proj": {
841
+ "bits": 8
842
+ },
843
+ "model.layers.27.self_attn.indexer.weights_proj": {
844
+ "bits": 16,
845
+ "data_type": "float"
846
+ },
847
+ "model.layers.27.self_attn.indexer.wk": {
848
+ "bits": 8
849
+ },
850
+ "model.layers.27.self_attn.indexer.wq_b": {
851
+ "bits": 8
852
+ },
853
+ "model.layers.27.self_attn.kv_a_proj_with_mqa": {
854
+ "bits": 8
855
+ },
856
+ "model.layers.27.self_attn.kv_b_proj": {
857
+ "bits": 8
858
+ },
859
+ "model.layers.27.self_attn.o_proj": {
860
+ "bits": 8
861
+ },
862
+ "model.layers.27.self_attn.q_a_proj": {
863
+ "bits": 8
864
+ },
865
+ "model.layers.27.self_attn.q_b_proj": {
866
+ "bits": 8
867
+ },
868
+ "model.layers.28.mlp.shared_experts.down_proj": {
869
+ "bits": 8
870
+ },
871
+ "model.layers.28.mlp.shared_experts.gate_proj": {
872
+ "bits": 8
873
+ },
874
+ "model.layers.28.mlp.shared_experts.up_proj": {
875
+ "bits": 8
876
+ },
877
+ "model.layers.28.self_attn.indexer.weights_proj": {
878
+ "bits": 16,
879
+ "data_type": "float"
880
+ },
881
+ "model.layers.28.self_attn.indexer.wk": {
882
+ "bits": 8
883
+ },
884
+ "model.layers.28.self_attn.indexer.wq_b": {
885
+ "bits": 8
886
+ },
887
+ "model.layers.28.self_attn.kv_a_proj_with_mqa": {
888
+ "bits": 8
889
+ },
890
+ "model.layers.28.self_attn.kv_b_proj": {
891
+ "bits": 8
892
+ },
893
+ "model.layers.28.self_attn.o_proj": {
894
+ "bits": 8
895
+ },
896
+ "model.layers.28.self_attn.q_a_proj": {
897
+ "bits": 8
898
+ },
899
+ "model.layers.28.self_attn.q_b_proj": {
900
+ "bits": 8
901
+ },
902
+ "model.layers.29.mlp.shared_experts.down_proj": {
903
+ "bits": 8
904
+ },
905
+ "model.layers.29.mlp.shared_experts.gate_proj": {
906
+ "bits": 8
907
+ },
908
+ "model.layers.29.mlp.shared_experts.up_proj": {
909
+ "bits": 8
910
+ },
911
+ "model.layers.29.self_attn.indexer.weights_proj": {
912
+ "bits": 16,
913
+ "data_type": "float"
914
+ },
915
+ "model.layers.29.self_attn.indexer.wk": {
916
+ "bits": 8
917
+ },
918
+ "model.layers.29.self_attn.indexer.wq_b": {
919
+ "bits": 8
920
+ },
921
+ "model.layers.29.self_attn.kv_a_proj_with_mqa": {
922
+ "bits": 8
923
+ },
924
+ "model.layers.29.self_attn.kv_b_proj": {
925
+ "bits": 8
926
+ },
927
+ "model.layers.29.self_attn.o_proj": {
928
+ "bits": 8
929
+ },
930
+ "model.layers.29.self_attn.q_a_proj": {
931
+ "bits": 8
932
+ },
933
+ "model.layers.29.self_attn.q_b_proj": {
934
+ "bits": 8
935
+ },
936
+ "model.layers.3.mlp.shared_experts.down_proj": {
937
+ "bits": 8
938
+ },
939
+ "model.layers.3.mlp.shared_experts.gate_proj": {
940
+ "bits": 8
941
+ },
942
+ "model.layers.3.mlp.shared_experts.up_proj": {
943
+ "bits": 8
944
+ },
945
+ "model.layers.3.self_attn.indexer.weights_proj": {
946
+ "bits": 16,
947
+ "data_type": "float"
948
+ },
949
+ "model.layers.3.self_attn.indexer.wk": {
950
+ "bits": 8
951
+ },
952
+ "model.layers.3.self_attn.indexer.wq_b": {
953
+ "bits": 8
954
+ },
955
+ "model.layers.3.self_attn.kv_a_proj_with_mqa": {
956
+ "bits": 8
957
+ },
958
+ "model.layers.3.self_attn.kv_b_proj": {
959
+ "bits": 8
960
+ },
961
+ "model.layers.3.self_attn.o_proj": {
962
+ "bits": 8
963
+ },
964
+ "model.layers.3.self_attn.q_a_proj": {
965
+ "bits": 8
966
+ },
967
+ "model.layers.3.self_attn.q_b_proj": {
968
+ "bits": 8
969
+ },
970
+ "model.layers.30.mlp.shared_experts.down_proj": {
971
+ "bits": 8
972
+ },
973
+ "model.layers.30.mlp.shared_experts.gate_proj": {
974
+ "bits": 8
975
+ },
976
+ "model.layers.30.mlp.shared_experts.up_proj": {
977
+ "bits": 8
978
+ },
979
+ "model.layers.30.self_attn.indexer.weights_proj": {
980
+ "bits": 16,
981
+ "data_type": "float"
982
+ },
983
+ "model.layers.30.self_attn.indexer.wk": {
984
+ "bits": 8
985
+ },
986
+ "model.layers.30.self_attn.indexer.wq_b": {
987
+ "bits": 8
988
+ },
989
+ "model.layers.30.self_attn.kv_a_proj_with_mqa": {
990
+ "bits": 8
991
+ },
992
+ "model.layers.30.self_attn.kv_b_proj": {
993
+ "bits": 8
994
+ },
995
+ "model.layers.30.self_attn.o_proj": {
996
+ "bits": 8
997
+ },
998
+ "model.layers.30.self_attn.q_a_proj": {
999
+ "bits": 8
1000
+ },
1001
+ "model.layers.30.self_attn.q_b_proj": {
1002
+ "bits": 8
1003
+ },
1004
+ "model.layers.31.mlp.shared_experts.down_proj": {
1005
+ "bits": 8
1006
+ },
1007
+ "model.layers.31.mlp.shared_experts.gate_proj": {
1008
+ "bits": 8
1009
+ },
1010
+ "model.layers.31.mlp.shared_experts.up_proj": {
1011
+ "bits": 8
1012
+ },
1013
+ "model.layers.31.self_attn.indexer.weights_proj": {
1014
+ "bits": 16,
1015
+ "data_type": "float"
1016
+ },
1017
+ "model.layers.31.self_attn.indexer.wk": {
1018
+ "bits": 8
1019
+ },
1020
+ "model.layers.31.self_attn.indexer.wq_b": {
1021
+ "bits": 8
1022
+ },
1023
+ "model.layers.31.self_attn.kv_a_proj_with_mqa": {
1024
+ "bits": 8
1025
+ },
1026
+ "model.layers.31.self_attn.kv_b_proj": {
1027
+ "bits": 8
1028
+ },
1029
+ "model.layers.31.self_attn.o_proj": {
1030
+ "bits": 8
1031
+ },
1032
+ "model.layers.31.self_attn.q_a_proj": {
1033
+ "bits": 8
1034
+ },
1035
+ "model.layers.31.self_attn.q_b_proj": {
1036
+ "bits": 8
1037
+ },
1038
+ "model.layers.32.mlp.shared_experts.down_proj": {
1039
+ "bits": 8
1040
+ },
1041
+ "model.layers.32.mlp.shared_experts.gate_proj": {
1042
+ "bits": 8
1043
+ },
1044
+ "model.layers.32.mlp.shared_experts.up_proj": {
1045
+ "bits": 8
1046
+ },
1047
+ "model.layers.32.self_attn.indexer.weights_proj": {
1048
+ "bits": 16,
1049
+ "data_type": "float"
1050
+ },
1051
+ "model.layers.32.self_attn.indexer.wk": {
1052
+ "bits": 8
1053
+ },
1054
+ "model.layers.32.self_attn.indexer.wq_b": {
1055
+ "bits": 8
1056
+ },
1057
+ "model.layers.32.self_attn.kv_a_proj_with_mqa": {
1058
+ "bits": 8
1059
+ },
1060
+ "model.layers.32.self_attn.kv_b_proj": {
1061
+ "bits": 8
1062
+ },
1063
+ "model.layers.32.self_attn.o_proj": {
1064
+ "bits": 8
1065
+ },
1066
+ "model.layers.32.self_attn.q_a_proj": {
1067
+ "bits": 8
1068
+ },
1069
+ "model.layers.32.self_attn.q_b_proj": {
1070
+ "bits": 8
1071
+ },
1072
+ "model.layers.33.mlp.shared_experts.down_proj": {
1073
+ "bits": 8
1074
+ },
1075
+ "model.layers.33.mlp.shared_experts.gate_proj": {
1076
+ "bits": 8
1077
+ },
1078
+ "model.layers.33.mlp.shared_experts.up_proj": {
1079
+ "bits": 8
1080
+ },
1081
+ "model.layers.33.self_attn.indexer.weights_proj": {
1082
+ "bits": 16,
1083
+ "data_type": "float"
1084
+ },
1085
+ "model.layers.33.self_attn.indexer.wk": {
1086
+ "bits": 8
1087
+ },
1088
+ "model.layers.33.self_attn.indexer.wq_b": {
1089
+ "bits": 8
1090
+ },
1091
+ "model.layers.33.self_attn.kv_a_proj_with_mqa": {
1092
+ "bits": 8
1093
+ },
1094
+ "model.layers.33.self_attn.kv_b_proj": {
1095
+ "bits": 8
1096
+ },
1097
+ "model.layers.33.self_attn.o_proj": {
1098
+ "bits": 8
1099
+ },
1100
+ "model.layers.33.self_attn.q_a_proj": {
1101
+ "bits": 8
1102
+ },
1103
+ "model.layers.33.self_attn.q_b_proj": {
1104
+ "bits": 8
1105
+ },
1106
+ "model.layers.34.mlp.shared_experts.down_proj": {
1107
+ "bits": 8
1108
+ },
1109
+ "model.layers.34.mlp.shared_experts.gate_proj": {
1110
+ "bits": 8
1111
+ },
1112
+ "model.layers.34.mlp.shared_experts.up_proj": {
1113
+ "bits": 8
1114
+ },
1115
+ "model.layers.34.self_attn.indexer.weights_proj": {
1116
+ "bits": 16,
1117
+ "data_type": "float"
1118
+ },
1119
+ "model.layers.34.self_attn.indexer.wk": {
1120
+ "bits": 8
1121
+ },
1122
+ "model.layers.34.self_attn.indexer.wq_b": {
1123
+ "bits": 8
1124
+ },
1125
+ "model.layers.34.self_attn.kv_a_proj_with_mqa": {
1126
+ "bits": 8
1127
+ },
1128
+ "model.layers.34.self_attn.kv_b_proj": {
1129
+ "bits": 8
1130
+ },
1131
+ "model.layers.34.self_attn.o_proj": {
1132
+ "bits": 8
1133
+ },
1134
+ "model.layers.34.self_attn.q_a_proj": {
1135
+ "bits": 8
1136
+ },
1137
+ "model.layers.34.self_attn.q_b_proj": {
1138
+ "bits": 8
1139
+ },
1140
+ "model.layers.35.mlp.shared_experts.down_proj": {
1141
+ "bits": 8
1142
+ },
1143
+ "model.layers.35.mlp.shared_experts.gate_proj": {
1144
+ "bits": 8
1145
+ },
1146
+ "model.layers.35.mlp.shared_experts.up_proj": {
1147
+ "bits": 8
1148
+ },
1149
+ "model.layers.35.self_attn.indexer.weights_proj": {
1150
+ "bits": 16,
1151
+ "data_type": "float"
1152
+ },
1153
+ "model.layers.35.self_attn.indexer.wk": {
1154
+ "bits": 8
1155
+ },
1156
+ "model.layers.35.self_attn.indexer.wq_b": {
1157
+ "bits": 8
1158
+ },
1159
+ "model.layers.35.self_attn.kv_a_proj_with_mqa": {
1160
+ "bits": 8
1161
+ },
1162
+ "model.layers.35.self_attn.kv_b_proj": {
1163
+ "bits": 8
1164
+ },
1165
+ "model.layers.35.self_attn.o_proj": {
1166
+ "bits": 8
1167
+ },
1168
+ "model.layers.35.self_attn.q_a_proj": {
1169
+ "bits": 8
1170
+ },
1171
+ "model.layers.35.self_attn.q_b_proj": {
1172
+ "bits": 8
1173
+ },
1174
+ "model.layers.36.mlp.shared_experts.down_proj": {
1175
+ "bits": 8
1176
+ },
1177
+ "model.layers.36.mlp.shared_experts.gate_proj": {
1178
+ "bits": 8
1179
+ },
1180
+ "model.layers.36.mlp.shared_experts.up_proj": {
1181
+ "bits": 8
1182
+ },
1183
+ "model.layers.36.self_attn.indexer.weights_proj": {
1184
+ "bits": 16,
1185
+ "data_type": "float"
1186
+ },
1187
+ "model.layers.36.self_attn.indexer.wk": {
1188
+ "bits": 8
1189
+ },
1190
+ "model.layers.36.self_attn.indexer.wq_b": {
1191
+ "bits": 8
1192
+ },
1193
+ "model.layers.36.self_attn.kv_a_proj_with_mqa": {
1194
+ "bits": 8
1195
+ },
1196
+ "model.layers.36.self_attn.kv_b_proj": {
1197
+ "bits": 8
1198
+ },
1199
+ "model.layers.36.self_attn.o_proj": {
1200
+ "bits": 8
1201
+ },
1202
+ "model.layers.36.self_attn.q_a_proj": {
1203
+ "bits": 8
1204
+ },
1205
+ "model.layers.36.self_attn.q_b_proj": {
1206
+ "bits": 8
1207
+ },
1208
+ "model.layers.37.mlp.shared_experts.down_proj": {
1209
+ "bits": 8
1210
+ },
1211
+ "model.layers.37.mlp.shared_experts.gate_proj": {
1212
+ "bits": 8
1213
+ },
1214
+ "model.layers.37.mlp.shared_experts.up_proj": {
1215
+ "bits": 8
1216
+ },
1217
+ "model.layers.37.self_attn.indexer.weights_proj": {
1218
+ "bits": 16,
1219
+ "data_type": "float"
1220
+ },
1221
+ "model.layers.37.self_attn.indexer.wk": {
1222
+ "bits": 8
1223
+ },
1224
+ "model.layers.37.self_attn.indexer.wq_b": {
1225
+ "bits": 8
1226
+ },
1227
+ "model.layers.37.self_attn.kv_a_proj_with_mqa": {
1228
+ "bits": 8
1229
+ },
1230
+ "model.layers.37.self_attn.kv_b_proj": {
1231
+ "bits": 8
1232
+ },
1233
+ "model.layers.37.self_attn.o_proj": {
1234
+ "bits": 8
1235
+ },
1236
+ "model.layers.37.self_attn.q_a_proj": {
1237
+ "bits": 8
1238
+ },
1239
+ "model.layers.37.self_attn.q_b_proj": {
1240
+ "bits": 8
1241
+ },
1242
+ "model.layers.38.mlp.shared_experts.down_proj": {
1243
+ "bits": 8
1244
+ },
1245
+ "model.layers.38.mlp.shared_experts.gate_proj": {
1246
+ "bits": 8
1247
+ },
1248
+ "model.layers.38.mlp.shared_experts.up_proj": {
1249
+ "bits": 8
1250
+ },
1251
+ "model.layers.38.self_attn.indexer.weights_proj": {
1252
+ "bits": 16,
1253
+ "data_type": "float"
1254
+ },
1255
+ "model.layers.38.self_attn.indexer.wk": {
1256
+ "bits": 8
1257
+ },
1258
+ "model.layers.38.self_attn.indexer.wq_b": {
1259
+ "bits": 8
1260
+ },
1261
+ "model.layers.38.self_attn.kv_a_proj_with_mqa": {
1262
+ "bits": 8
1263
+ },
1264
+ "model.layers.38.self_attn.kv_b_proj": {
1265
+ "bits": 8
1266
+ },
1267
+ "model.layers.38.self_attn.o_proj": {
1268
+ "bits": 8
1269
+ },
1270
+ "model.layers.38.self_attn.q_a_proj": {
1271
+ "bits": 8
1272
+ },
1273
+ "model.layers.38.self_attn.q_b_proj": {
1274
+ "bits": 8
1275
+ },
1276
+ "model.layers.39.mlp.shared_experts.down_proj": {
1277
+ "bits": 8
1278
+ },
1279
+ "model.layers.39.mlp.shared_experts.gate_proj": {
1280
+ "bits": 8
1281
+ },
1282
+ "model.layers.39.mlp.shared_experts.up_proj": {
1283
+ "bits": 8
1284
+ },
1285
+ "model.layers.39.self_attn.indexer.weights_proj": {
1286
+ "bits": 16,
1287
+ "data_type": "float"
1288
+ },
1289
+ "model.layers.39.self_attn.indexer.wk": {
1290
+ "bits": 8
1291
+ },
1292
+ "model.layers.39.self_attn.indexer.wq_b": {
1293
+ "bits": 8
1294
+ },
1295
+ "model.layers.39.self_attn.kv_a_proj_with_mqa": {
1296
+ "bits": 8
1297
+ },
1298
+ "model.layers.39.self_attn.kv_b_proj": {
1299
+ "bits": 8
1300
+ },
1301
+ "model.layers.39.self_attn.o_proj": {
1302
+ "bits": 8
1303
+ },
1304
+ "model.layers.39.self_attn.q_a_proj": {
1305
+ "bits": 8
1306
+ },
1307
+ "model.layers.39.self_attn.q_b_proj": {
1308
+ "bits": 8
1309
+ },
1310
+ "model.layers.4.mlp.shared_experts.down_proj": {
1311
+ "bits": 8
1312
+ },
1313
+ "model.layers.4.mlp.shared_experts.gate_proj": {
1314
+ "bits": 8
1315
+ },
1316
+ "model.layers.4.mlp.shared_experts.up_proj": {
1317
+ "bits": 8
1318
+ },
1319
+ "model.layers.4.self_attn.indexer.weights_proj": {
1320
+ "bits": 16,
1321
+ "data_type": "float"
1322
+ },
1323
+ "model.layers.4.self_attn.indexer.wk": {
1324
+ "bits": 8
1325
+ },
1326
+ "model.layers.4.self_attn.indexer.wq_b": {
1327
+ "bits": 8
1328
+ },
1329
+ "model.layers.4.self_attn.kv_a_proj_with_mqa": {
1330
+ "bits": 8
1331
+ },
1332
+ "model.layers.4.self_attn.kv_b_proj": {
1333
+ "bits": 8
1334
+ },
1335
+ "model.layers.4.self_attn.o_proj": {
1336
+ "bits": 8
1337
+ },
1338
+ "model.layers.4.self_attn.q_a_proj": {
1339
+ "bits": 8
1340
+ },
1341
+ "model.layers.4.self_attn.q_b_proj": {
1342
+ "bits": 8
1343
+ },
1344
+ "model.layers.40.mlp.shared_experts.down_proj": {
1345
+ "bits": 8
1346
+ },
1347
+ "model.layers.40.mlp.shared_experts.gate_proj": {
1348
+ "bits": 8
1349
+ },
1350
+ "model.layers.40.mlp.shared_experts.up_proj": {
1351
+ "bits": 8
1352
+ },
1353
+ "model.layers.40.self_attn.indexer.weights_proj": {
1354
+ "bits": 16,
1355
+ "data_type": "float"
1356
+ },
1357
+ "model.layers.40.self_attn.indexer.wk": {
1358
+ "bits": 8
1359
+ },
1360
+ "model.layers.40.self_attn.indexer.wq_b": {
1361
+ "bits": 8
1362
+ },
1363
+ "model.layers.40.self_attn.kv_a_proj_with_mqa": {
1364
+ "bits": 8
1365
+ },
1366
+ "model.layers.40.self_attn.kv_b_proj": {
1367
+ "bits": 8
1368
+ },
1369
+ "model.layers.40.self_attn.o_proj": {
1370
+ "bits": 8
1371
+ },
1372
+ "model.layers.40.self_attn.q_a_proj": {
1373
+ "bits": 8
1374
+ },
1375
+ "model.layers.40.self_attn.q_b_proj": {
1376
+ "bits": 8
1377
+ },
1378
+ "model.layers.41.mlp.shared_experts.down_proj": {
1379
+ "bits": 8
1380
+ },
1381
+ "model.layers.41.mlp.shared_experts.gate_proj": {
1382
+ "bits": 8
1383
+ },
1384
+ "model.layers.41.mlp.shared_experts.up_proj": {
1385
+ "bits": 8
1386
+ },
1387
+ "model.layers.41.self_attn.indexer.weights_proj": {
1388
+ "bits": 16,
1389
+ "data_type": "float"
1390
+ },
1391
+ "model.layers.41.self_attn.indexer.wk": {
1392
+ "bits": 8
1393
+ },
1394
+ "model.layers.41.self_attn.indexer.wq_b": {
1395
+ "bits": 8
1396
+ },
1397
+ "model.layers.41.self_attn.kv_a_proj_with_mqa": {
1398
+ "bits": 8
1399
+ },
1400
+ "model.layers.41.self_attn.kv_b_proj": {
1401
+ "bits": 8
1402
+ },
1403
+ "model.layers.41.self_attn.o_proj": {
1404
+ "bits": 8
1405
+ },
1406
+ "model.layers.41.self_attn.q_a_proj": {
1407
+ "bits": 8
1408
+ },
1409
+ "model.layers.41.self_attn.q_b_proj": {
1410
+ "bits": 8
1411
+ },
1412
+ "model.layers.42.mlp.shared_experts.down_proj": {
1413
+ "bits": 8
1414
+ },
1415
+ "model.layers.42.mlp.shared_experts.gate_proj": {
1416
+ "bits": 8
1417
+ },
1418
+ "model.layers.42.mlp.shared_experts.up_proj": {
1419
+ "bits": 8
1420
+ },
1421
+ "model.layers.42.self_attn.indexer.weights_proj": {
1422
+ "bits": 16,
1423
+ "data_type": "float"
1424
+ },
1425
+ "model.layers.42.self_attn.indexer.wk": {
1426
+ "bits": 8
1427
+ },
1428
+ "model.layers.42.self_attn.indexer.wq_b": {
1429
+ "bits": 8
1430
+ },
1431
+ "model.layers.42.self_attn.kv_a_proj_with_mqa": {
1432
+ "bits": 8
1433
+ },
1434
+ "model.layers.42.self_attn.kv_b_proj": {
1435
+ "bits": 8
1436
+ },
1437
+ "model.layers.42.self_attn.o_proj": {
1438
+ "bits": 8
1439
+ },
1440
+ "model.layers.42.self_attn.q_a_proj": {
1441
+ "bits": 8
1442
+ },
1443
+ "model.layers.42.self_attn.q_b_proj": {
1444
+ "bits": 8
1445
+ },
1446
+ "model.layers.43.mlp.shared_experts.down_proj": {
1447
+ "bits": 8
1448
+ },
1449
+ "model.layers.43.mlp.shared_experts.gate_proj": {
1450
+ "bits": 8
1451
+ },
1452
+ "model.layers.43.mlp.shared_experts.up_proj": {
1453
+ "bits": 8
1454
+ },
1455
+ "model.layers.43.self_attn.indexer.weights_proj": {
1456
+ "bits": 16,
1457
+ "data_type": "float"
1458
+ },
1459
+ "model.layers.43.self_attn.indexer.wk": {
1460
+ "bits": 8
1461
+ },
1462
+ "model.layers.43.self_attn.indexer.wq_b": {
1463
+ "bits": 8
1464
+ },
1465
+ "model.layers.43.self_attn.kv_a_proj_with_mqa": {
1466
+ "bits": 8
1467
+ },
1468
+ "model.layers.43.self_attn.kv_b_proj": {
1469
+ "bits": 8
1470
+ },
1471
+ "model.layers.43.self_attn.o_proj": {
1472
+ "bits": 8
1473
+ },
1474
+ "model.layers.43.self_attn.q_a_proj": {
1475
+ "bits": 8
1476
+ },
1477
+ "model.layers.43.self_attn.q_b_proj": {
1478
+ "bits": 8
1479
+ },
1480
+ "model.layers.44.mlp.shared_experts.down_proj": {
1481
+ "bits": 8
1482
+ },
1483
+ "model.layers.44.mlp.shared_experts.gate_proj": {
1484
+ "bits": 8
1485
+ },
1486
+ "model.layers.44.mlp.shared_experts.up_proj": {
1487
+ "bits": 8
1488
+ },
1489
+ "model.layers.44.self_attn.indexer.weights_proj": {
1490
+ "bits": 16,
1491
+ "data_type": "float"
1492
+ },
1493
+ "model.layers.44.self_attn.indexer.wk": {
1494
+ "bits": 8
1495
+ },
1496
+ "model.layers.44.self_attn.indexer.wq_b": {
1497
+ "bits": 8
1498
+ },
1499
+ "model.layers.44.self_attn.kv_a_proj_with_mqa": {
1500
+ "bits": 8
1501
+ },
1502
+ "model.layers.44.self_attn.kv_b_proj": {
1503
+ "bits": 8
1504
+ },
1505
+ "model.layers.44.self_attn.o_proj": {
1506
+ "bits": 8
1507
+ },
1508
+ "model.layers.44.self_attn.q_a_proj": {
1509
+ "bits": 8
1510
+ },
1511
+ "model.layers.44.self_attn.q_b_proj": {
1512
+ "bits": 8
1513
+ },
1514
+ "model.layers.45.mlp.shared_experts.down_proj": {
1515
+ "bits": 8
1516
+ },
1517
+ "model.layers.45.mlp.shared_experts.gate_proj": {
1518
+ "bits": 8
1519
+ },
1520
+ "model.layers.45.mlp.shared_experts.up_proj": {
1521
+ "bits": 8
1522
+ },
1523
+ "model.layers.45.self_attn.indexer.weights_proj": {
1524
+ "bits": 16,
1525
+ "data_type": "float"
1526
+ },
1527
+ "model.layers.45.self_attn.indexer.wk": {
1528
+ "bits": 8
1529
+ },
1530
+ "model.layers.45.self_attn.indexer.wq_b": {
1531
+ "bits": 8
1532
+ },
1533
+ "model.layers.45.self_attn.kv_a_proj_with_mqa": {
1534
+ "bits": 8
1535
+ },
1536
+ "model.layers.45.self_attn.kv_b_proj": {
1537
+ "bits": 8
1538
+ },
1539
+ "model.layers.45.self_attn.o_proj": {
1540
+ "bits": 8
1541
+ },
1542
+ "model.layers.45.self_attn.q_a_proj": {
1543
+ "bits": 8
1544
+ },
1545
+ "model.layers.45.self_attn.q_b_proj": {
1546
+ "bits": 8
1547
+ },
1548
+ "model.layers.46.mlp.shared_experts.down_proj": {
1549
+ "bits": 8
1550
+ },
1551
+ "model.layers.46.mlp.shared_experts.gate_proj": {
1552
+ "bits": 8
1553
+ },
1554
+ "model.layers.46.mlp.shared_experts.up_proj": {
1555
+ "bits": 8
1556
+ },
1557
+ "model.layers.46.self_attn.indexer.weights_proj": {
1558
+ "bits": 16,
1559
+ "data_type": "float"
1560
+ },
1561
+ "model.layers.46.self_attn.indexer.wk": {
1562
+ "bits": 8
1563
+ },
1564
+ "model.layers.46.self_attn.indexer.wq_b": {
1565
+ "bits": 8
1566
+ },
1567
+ "model.layers.46.self_attn.kv_a_proj_with_mqa": {
1568
+ "bits": 8
1569
+ },
1570
+ "model.layers.46.self_attn.kv_b_proj": {
1571
+ "bits": 8
1572
+ },
1573
+ "model.layers.46.self_attn.o_proj": {
1574
+ "bits": 8
1575
+ },
1576
+ "model.layers.46.self_attn.q_a_proj": {
1577
+ "bits": 8
1578
+ },
1579
+ "model.layers.46.self_attn.q_b_proj": {
1580
+ "bits": 8
1581
+ },
1582
+ "model.layers.47.mlp.shared_experts.down_proj": {
1583
+ "bits": 8
1584
+ },
1585
+ "model.layers.47.mlp.shared_experts.gate_proj": {
1586
+ "bits": 8
1587
+ },
1588
+ "model.layers.47.mlp.shared_experts.up_proj": {
1589
+ "bits": 8
1590
+ },
1591
+ "model.layers.47.self_attn.indexer.weights_proj": {
1592
+ "bits": 16,
1593
+ "data_type": "float"
1594
+ },
1595
+ "model.layers.47.self_attn.indexer.wk": {
1596
+ "bits": 8
1597
+ },
1598
+ "model.layers.47.self_attn.indexer.wq_b": {
1599
+ "bits": 8
1600
+ },
1601
+ "model.layers.47.self_attn.kv_a_proj_with_mqa": {
1602
+ "bits": 8
1603
+ },
1604
+ "model.layers.47.self_attn.kv_b_proj": {
1605
+ "bits": 8
1606
+ },
1607
+ "model.layers.47.self_attn.o_proj": {
1608
+ "bits": 8
1609
+ },
1610
+ "model.layers.47.self_attn.q_a_proj": {
1611
+ "bits": 8
1612
+ },
1613
+ "model.layers.47.self_attn.q_b_proj": {
1614
+ "bits": 8
1615
+ },
1616
+ "model.layers.48.mlp.shared_experts.down_proj": {
1617
+ "bits": 8
1618
+ },
1619
+ "model.layers.48.mlp.shared_experts.gate_proj": {
1620
+ "bits": 8
1621
+ },
1622
+ "model.layers.48.mlp.shared_experts.up_proj": {
1623
+ "bits": 8
1624
+ },
1625
+ "model.layers.48.self_attn.indexer.weights_proj": {
1626
+ "bits": 16,
1627
+ "data_type": "float"
1628
+ },
1629
+ "model.layers.48.self_attn.indexer.wk": {
1630
+ "bits": 8
1631
+ },
1632
+ "model.layers.48.self_attn.indexer.wq_b": {
1633
+ "bits": 8
1634
+ },
1635
+ "model.layers.48.self_attn.kv_a_proj_with_mqa": {
1636
+ "bits": 8
1637
+ },
1638
+ "model.layers.48.self_attn.kv_b_proj": {
1639
+ "bits": 8
1640
+ },
1641
+ "model.layers.48.self_attn.o_proj": {
1642
+ "bits": 8
1643
+ },
1644
+ "model.layers.48.self_attn.q_a_proj": {
1645
+ "bits": 8
1646
+ },
1647
+ "model.layers.48.self_attn.q_b_proj": {
1648
+ "bits": 8
1649
+ },
1650
+ "model.layers.49.mlp.shared_experts.down_proj": {
1651
+ "bits": 8
1652
+ },
1653
+ "model.layers.49.mlp.shared_experts.gate_proj": {
1654
+ "bits": 8
1655
+ },
1656
+ "model.layers.49.mlp.shared_experts.up_proj": {
1657
+ "bits": 8
1658
+ },
1659
+ "model.layers.49.self_attn.indexer.weights_proj": {
1660
+ "bits": 16,
1661
+ "data_type": "float"
1662
+ },
1663
+ "model.layers.49.self_attn.indexer.wk": {
1664
+ "bits": 8
1665
+ },
1666
+ "model.layers.49.self_attn.indexer.wq_b": {
1667
+ "bits": 8
1668
+ },
1669
+ "model.layers.49.self_attn.kv_a_proj_with_mqa": {
1670
+ "bits": 8
1671
+ },
1672
+ "model.layers.49.self_attn.kv_b_proj": {
1673
+ "bits": 8
1674
+ },
1675
+ "model.layers.49.self_attn.o_proj": {
1676
+ "bits": 8
1677
+ },
1678
+ "model.layers.49.self_attn.q_a_proj": {
1679
+ "bits": 8
1680
+ },
1681
+ "model.layers.49.self_attn.q_b_proj": {
1682
+ "bits": 8
1683
+ },
1684
+ "model.layers.5.mlp.shared_experts.down_proj": {
1685
+ "bits": 8
1686
+ },
1687
+ "model.layers.5.mlp.shared_experts.gate_proj": {
1688
+ "bits": 8
1689
+ },
1690
+ "model.layers.5.mlp.shared_experts.up_proj": {
1691
+ "bits": 8
1692
+ },
1693
+ "model.layers.5.self_attn.indexer.weights_proj": {
1694
+ "bits": 16,
1695
+ "data_type": "float"
1696
+ },
1697
+ "model.layers.5.self_attn.indexer.wk": {
1698
+ "bits": 8
1699
+ },
1700
+ "model.layers.5.self_attn.indexer.wq_b": {
1701
+ "bits": 8
1702
+ },
1703
+ "model.layers.5.self_attn.kv_a_proj_with_mqa": {
1704
+ "bits": 8
1705
+ },
1706
+ "model.layers.5.self_attn.kv_b_proj": {
1707
+ "bits": 8
1708
+ },
1709
+ "model.layers.5.self_attn.o_proj": {
1710
+ "bits": 8
1711
+ },
1712
+ "model.layers.5.self_attn.q_a_proj": {
1713
+ "bits": 8
1714
+ },
1715
+ "model.layers.5.self_attn.q_b_proj": {
1716
+ "bits": 8
1717
+ },
1718
+ "model.layers.50.mlp.shared_experts.down_proj": {
1719
+ "bits": 8
1720
+ },
1721
+ "model.layers.50.mlp.shared_experts.gate_proj": {
1722
+ "bits": 8
1723
+ },
1724
+ "model.layers.50.mlp.shared_experts.up_proj": {
1725
+ "bits": 8
1726
+ },
1727
+ "model.layers.50.self_attn.indexer.weights_proj": {
1728
+ "bits": 16,
1729
+ "data_type": "float"
1730
+ },
1731
+ "model.layers.50.self_attn.indexer.wk": {
1732
+ "bits": 8
1733
+ },
1734
+ "model.layers.50.self_attn.indexer.wq_b": {
1735
+ "bits": 8
1736
+ },
1737
+ "model.layers.50.self_attn.kv_a_proj_with_mqa": {
1738
+ "bits": 8
1739
+ },
1740
+ "model.layers.50.self_attn.kv_b_proj": {
1741
+ "bits": 8
1742
+ },
1743
+ "model.layers.50.self_attn.o_proj": {
1744
+ "bits": 8
1745
+ },
1746
+ "model.layers.50.self_attn.q_a_proj": {
1747
+ "bits": 8
1748
+ },
1749
+ "model.layers.50.self_attn.q_b_proj": {
1750
+ "bits": 8
1751
+ },
1752
+ "model.layers.51.mlp.shared_experts.down_proj": {
1753
+ "bits": 8
1754
+ },
1755
+ "model.layers.51.mlp.shared_experts.gate_proj": {
1756
+ "bits": 8
1757
+ },
1758
+ "model.layers.51.mlp.shared_experts.up_proj": {
1759
+ "bits": 8
1760
+ },
1761
+ "model.layers.51.self_attn.indexer.weights_proj": {
1762
+ "bits": 16,
1763
+ "data_type": "float"
1764
+ },
1765
+ "model.layers.51.self_attn.indexer.wk": {
1766
+ "bits": 8
1767
+ },
1768
+ "model.layers.51.self_attn.indexer.wq_b": {
1769
+ "bits": 8
1770
+ },
1771
+ "model.layers.51.self_attn.kv_a_proj_with_mqa": {
1772
+ "bits": 8
1773
+ },
1774
+ "model.layers.51.self_attn.kv_b_proj": {
1775
+ "bits": 8
1776
+ },
1777
+ "model.layers.51.self_attn.o_proj": {
1778
+ "bits": 8
1779
+ },
1780
+ "model.layers.51.self_attn.q_a_proj": {
1781
+ "bits": 8
1782
+ },
1783
+ "model.layers.51.self_attn.q_b_proj": {
1784
+ "bits": 8
1785
+ },
1786
+ "model.layers.52.mlp.shared_experts.down_proj": {
1787
+ "bits": 8
1788
+ },
1789
+ "model.layers.52.mlp.shared_experts.gate_proj": {
1790
+ "bits": 8
1791
+ },
1792
+ "model.layers.52.mlp.shared_experts.up_proj": {
1793
+ "bits": 8
1794
+ },
1795
+ "model.layers.52.self_attn.indexer.weights_proj": {
1796
+ "bits": 16,
1797
+ "data_type": "float"
1798
+ },
1799
+ "model.layers.52.self_attn.indexer.wk": {
1800
+ "bits": 8
1801
+ },
1802
+ "model.layers.52.self_attn.indexer.wq_b": {
1803
+ "bits": 8
1804
+ },
1805
+ "model.layers.52.self_attn.kv_a_proj_with_mqa": {
1806
+ "bits": 8
1807
+ },
1808
+ "model.layers.52.self_attn.kv_b_proj": {
1809
+ "bits": 8
1810
+ },
1811
+ "model.layers.52.self_attn.o_proj": {
1812
+ "bits": 8
1813
+ },
1814
+ "model.layers.52.self_attn.q_a_proj": {
1815
+ "bits": 8
1816
+ },
1817
+ "model.layers.52.self_attn.q_b_proj": {
1818
+ "bits": 8
1819
+ },
1820
+ "model.layers.53.mlp.shared_experts.down_proj": {
1821
+ "bits": 8
1822
+ },
1823
+ "model.layers.53.mlp.shared_experts.gate_proj": {
1824
+ "bits": 8
1825
+ },
1826
+ "model.layers.53.mlp.shared_experts.up_proj": {
1827
+ "bits": 8
1828
+ },
1829
+ "model.layers.53.self_attn.indexer.weights_proj": {
1830
+ "bits": 16,
1831
+ "data_type": "float"
1832
+ },
1833
+ "model.layers.53.self_attn.indexer.wk": {
1834
+ "bits": 8
1835
+ },
1836
+ "model.layers.53.self_attn.indexer.wq_b": {
1837
+ "bits": 8
1838
+ },
1839
+ "model.layers.53.self_attn.kv_a_proj_with_mqa": {
1840
+ "bits": 8
1841
+ },
1842
+ "model.layers.53.self_attn.kv_b_proj": {
1843
+ "bits": 8
1844
+ },
1845
+ "model.layers.53.self_attn.o_proj": {
1846
+ "bits": 8
1847
+ },
1848
+ "model.layers.53.self_attn.q_a_proj": {
1849
+ "bits": 8
1850
+ },
1851
+ "model.layers.53.self_attn.q_b_proj": {
1852
+ "bits": 8
1853
+ },
1854
+ "model.layers.54.mlp.shared_experts.down_proj": {
1855
+ "bits": 8
1856
+ },
1857
+ "model.layers.54.mlp.shared_experts.gate_proj": {
1858
+ "bits": 8
1859
+ },
1860
+ "model.layers.54.mlp.shared_experts.up_proj": {
1861
+ "bits": 8
1862
+ },
1863
+ "model.layers.54.self_attn.indexer.weights_proj": {
1864
+ "bits": 16,
1865
+ "data_type": "float"
1866
+ },
1867
+ "model.layers.54.self_attn.indexer.wk": {
1868
+ "bits": 8
1869
+ },
1870
+ "model.layers.54.self_attn.indexer.wq_b": {
1871
+ "bits": 8
1872
+ },
1873
+ "model.layers.54.self_attn.kv_a_proj_with_mqa": {
1874
+ "bits": 8
1875
+ },
1876
+ "model.layers.54.self_attn.kv_b_proj": {
1877
+ "bits": 8
1878
+ },
1879
+ "model.layers.54.self_attn.o_proj": {
1880
+ "bits": 8
1881
+ },
1882
+ "model.layers.54.self_attn.q_a_proj": {
1883
+ "bits": 8
1884
+ },
1885
+ "model.layers.54.self_attn.q_b_proj": {
1886
+ "bits": 8
1887
+ },
1888
+ "model.layers.55.mlp.shared_experts.down_proj": {
1889
+ "bits": 8
1890
+ },
1891
+ "model.layers.55.mlp.shared_experts.gate_proj": {
1892
+ "bits": 8
1893
+ },
1894
+ "model.layers.55.mlp.shared_experts.up_proj": {
1895
+ "bits": 8
1896
+ },
1897
+ "model.layers.55.self_attn.indexer.weights_proj": {
1898
+ "bits": 16,
1899
+ "data_type": "float"
1900
+ },
1901
+ "model.layers.55.self_attn.indexer.wk": {
1902
+ "bits": 8
1903
+ },
1904
+ "model.layers.55.self_attn.indexer.wq_b": {
1905
+ "bits": 8
1906
+ },
1907
+ "model.layers.55.self_attn.kv_a_proj_with_mqa": {
1908
+ "bits": 8
1909
+ },
1910
+ "model.layers.55.self_attn.kv_b_proj": {
1911
+ "bits": 8
1912
+ },
1913
+ "model.layers.55.self_attn.o_proj": {
1914
+ "bits": 8
1915
+ },
1916
+ "model.layers.55.self_attn.q_a_proj": {
1917
+ "bits": 8
1918
+ },
1919
+ "model.layers.55.self_attn.q_b_proj": {
1920
+ "bits": 8
1921
+ },
1922
+ "model.layers.56.mlp.shared_experts.down_proj": {
1923
+ "bits": 8
1924
+ },
1925
+ "model.layers.56.mlp.shared_experts.gate_proj": {
1926
+ "bits": 8
1927
+ },
1928
+ "model.layers.56.mlp.shared_experts.up_proj": {
1929
+ "bits": 8
1930
+ },
1931
+ "model.layers.56.self_attn.indexer.weights_proj": {
1932
+ "bits": 16,
1933
+ "data_type": "float"
1934
+ },
1935
+ "model.layers.56.self_attn.indexer.wk": {
1936
+ "bits": 8
1937
+ },
1938
+ "model.layers.56.self_attn.indexer.wq_b": {
1939
+ "bits": 8
1940
+ },
1941
+ "model.layers.56.self_attn.kv_a_proj_with_mqa": {
1942
+ "bits": 8
1943
+ },
1944
+ "model.layers.56.self_attn.kv_b_proj": {
1945
+ "bits": 8
1946
+ },
1947
+ "model.layers.56.self_attn.o_proj": {
1948
+ "bits": 8
1949
+ },
1950
+ "model.layers.56.self_attn.q_a_proj": {
1951
+ "bits": 8
1952
+ },
1953
+ "model.layers.56.self_attn.q_b_proj": {
1954
+ "bits": 8
1955
+ },
1956
+ "model.layers.57.mlp.shared_experts.down_proj": {
1957
+ "bits": 8
1958
+ },
1959
+ "model.layers.57.mlp.shared_experts.gate_proj": {
1960
+ "bits": 8
1961
+ },
1962
+ "model.layers.57.mlp.shared_experts.up_proj": {
1963
+ "bits": 8
1964
+ },
1965
+ "model.layers.57.self_attn.indexer.weights_proj": {
1966
+ "bits": 16,
1967
+ "data_type": "float"
1968
+ },
1969
+ "model.layers.57.self_attn.indexer.wk": {
1970
+ "bits": 8
1971
+ },
1972
+ "model.layers.57.self_attn.indexer.wq_b": {
1973
+ "bits": 8
1974
+ },
1975
+ "model.layers.57.self_attn.kv_a_proj_with_mqa": {
1976
+ "bits": 8
1977
+ },
1978
+ "model.layers.57.self_attn.kv_b_proj": {
1979
+ "bits": 8
1980
+ },
1981
+ "model.layers.57.self_attn.o_proj": {
1982
+ "bits": 8
1983
+ },
1984
+ "model.layers.57.self_attn.q_a_proj": {
1985
+ "bits": 8
1986
+ },
1987
+ "model.layers.57.self_attn.q_b_proj": {
1988
+ "bits": 8
1989
+ },
1990
+ "model.layers.58.mlp.shared_experts.down_proj": {
1991
+ "bits": 8
1992
+ },
1993
+ "model.layers.58.mlp.shared_experts.gate_proj": {
1994
+ "bits": 8
1995
+ },
1996
+ "model.layers.58.mlp.shared_experts.up_proj": {
1997
+ "bits": 8
1998
+ },
1999
+ "model.layers.58.self_attn.indexer.weights_proj": {
2000
+ "bits": 16,
2001
+ "data_type": "float"
2002
+ },
2003
+ "model.layers.58.self_attn.indexer.wk": {
2004
+ "bits": 8
2005
+ },
2006
+ "model.layers.58.self_attn.indexer.wq_b": {
2007
+ "bits": 8
2008
+ },
2009
+ "model.layers.58.self_attn.kv_a_proj_with_mqa": {
2010
+ "bits": 8
2011
+ },
2012
+ "model.layers.58.self_attn.kv_b_proj": {
2013
+ "bits": 8
2014
+ },
2015
+ "model.layers.58.self_attn.o_proj": {
2016
+ "bits": 8
2017
+ },
2018
+ "model.layers.58.self_attn.q_a_proj": {
2019
+ "bits": 8
2020
+ },
2021
+ "model.layers.58.self_attn.q_b_proj": {
2022
+ "bits": 8
2023
+ },
2024
+ "model.layers.59.mlp.shared_experts.down_proj": {
2025
+ "bits": 8
2026
+ },
2027
+ "model.layers.59.mlp.shared_experts.gate_proj": {
2028
+ "bits": 8
2029
+ },
2030
+ "model.layers.59.mlp.shared_experts.up_proj": {
2031
+ "bits": 8
2032
+ },
2033
+ "model.layers.59.self_attn.indexer.weights_proj": {
2034
+ "bits": 16,
2035
+ "data_type": "float"
2036
+ },
2037
+ "model.layers.59.self_attn.indexer.wk": {
2038
+ "bits": 8
2039
+ },
2040
+ "model.layers.59.self_attn.indexer.wq_b": {
2041
+ "bits": 8
2042
+ },
2043
+ "model.layers.59.self_attn.kv_a_proj_with_mqa": {
2044
+ "bits": 8
2045
+ },
2046
+ "model.layers.59.self_attn.kv_b_proj": {
2047
+ "bits": 8
2048
+ },
2049
+ "model.layers.59.self_attn.o_proj": {
2050
+ "bits": 8
2051
+ },
2052
+ "model.layers.59.self_attn.q_a_proj": {
2053
+ "bits": 8
2054
+ },
2055
+ "model.layers.59.self_attn.q_b_proj": {
2056
+ "bits": 8
2057
+ },
2058
+ "model.layers.6.mlp.shared_experts.down_proj": {
2059
+ "bits": 8
2060
+ },
2061
+ "model.layers.6.mlp.shared_experts.gate_proj": {
2062
+ "bits": 8
2063
+ },
2064
+ "model.layers.6.mlp.shared_experts.up_proj": {
2065
+ "bits": 8
2066
+ },
2067
+ "model.layers.6.self_attn.indexer.weights_proj": {
2068
+ "bits": 16,
2069
+ "data_type": "float"
2070
+ },
2071
+ "model.layers.6.self_attn.indexer.wk": {
2072
+ "bits": 8
2073
+ },
2074
+ "model.layers.6.self_attn.indexer.wq_b": {
2075
+ "bits": 8
2076
+ },
2077
+ "model.layers.6.self_attn.kv_a_proj_with_mqa": {
2078
+ "bits": 8
2079
+ },
2080
+ "model.layers.6.self_attn.kv_b_proj": {
2081
+ "bits": 8
2082
+ },
2083
+ "model.layers.6.self_attn.o_proj": {
2084
+ "bits": 8
2085
+ },
2086
+ "model.layers.6.self_attn.q_a_proj": {
2087
+ "bits": 8
2088
+ },
2089
+ "model.layers.6.self_attn.q_b_proj": {
2090
+ "bits": 8
2091
+ },
2092
+ "model.layers.60.mlp.shared_experts.down_proj": {
2093
+ "bits": 8
2094
+ },
2095
+ "model.layers.60.mlp.shared_experts.gate_proj": {
2096
+ "bits": 8
2097
+ },
2098
+ "model.layers.60.mlp.shared_experts.up_proj": {
2099
+ "bits": 8
2100
+ },
2101
+ "model.layers.60.self_attn.indexer.weights_proj": {
2102
+ "bits": 16,
2103
+ "data_type": "float"
2104
+ },
2105
+ "model.layers.60.self_attn.indexer.wk": {
2106
+ "bits": 8
2107
+ },
2108
+ "model.layers.60.self_attn.indexer.wq_b": {
2109
+ "bits": 8
2110
+ },
2111
+ "model.layers.60.self_attn.kv_a_proj_with_mqa": {
2112
+ "bits": 8
2113
+ },
2114
+ "model.layers.60.self_attn.kv_b_proj": {
2115
+ "bits": 8
2116
+ },
2117
+ "model.layers.60.self_attn.o_proj": {
2118
+ "bits": 8
2119
+ },
2120
+ "model.layers.60.self_attn.q_a_proj": {
2121
+ "bits": 8
2122
+ },
2123
+ "model.layers.60.self_attn.q_b_proj": {
2124
+ "bits": 8
2125
+ },
2126
+ "model.layers.61.mlp.shared_experts.down_proj": {
2127
+ "bits": 8
2128
+ },
2129
+ "model.layers.61.mlp.shared_experts.gate_proj": {
2130
+ "bits": 8
2131
+ },
2132
+ "model.layers.61.mlp.shared_experts.up_proj": {
2133
+ "bits": 8
2134
+ },
2135
+ "model.layers.61.self_attn.indexer.weights_proj": {
2136
+ "bits": 16,
2137
+ "data_type": "float"
2138
+ },
2139
+ "model.layers.61.self_attn.indexer.wk": {
2140
+ "bits": 8
2141
+ },
2142
+ "model.layers.61.self_attn.indexer.wq_b": {
2143
+ "bits": 8
2144
+ },
2145
+ "model.layers.61.self_attn.kv_a_proj_with_mqa": {
2146
+ "bits": 8
2147
+ },
2148
+ "model.layers.61.self_attn.kv_b_proj": {
2149
+ "bits": 8
2150
+ },
2151
+ "model.layers.61.self_attn.o_proj": {
2152
+ "bits": 8
2153
+ },
2154
+ "model.layers.61.self_attn.q_a_proj": {
2155
+ "bits": 8
2156
+ },
2157
+ "model.layers.61.self_attn.q_b_proj": {
2158
+ "bits": 8
2159
+ },
2160
+ "model.layers.62.mlp.shared_experts.down_proj": {
2161
+ "bits": 8
2162
+ },
2163
+ "model.layers.62.mlp.shared_experts.gate_proj": {
2164
+ "bits": 8
2165
+ },
2166
+ "model.layers.62.mlp.shared_experts.up_proj": {
2167
+ "bits": 8
2168
+ },
2169
+ "model.layers.62.self_attn.indexer.weights_proj": {
2170
+ "bits": 16,
2171
+ "data_type": "float"
2172
+ },
2173
+ "model.layers.62.self_attn.indexer.wk": {
2174
+ "bits": 8
2175
+ },
2176
+ "model.layers.62.self_attn.indexer.wq_b": {
2177
+ "bits": 8
2178
+ },
2179
+ "model.layers.62.self_attn.kv_a_proj_with_mqa": {
2180
+ "bits": 8
2181
+ },
2182
+ "model.layers.62.self_attn.kv_b_proj": {
2183
+ "bits": 8
2184
+ },
2185
+ "model.layers.62.self_attn.o_proj": {
2186
+ "bits": 8
2187
+ },
2188
+ "model.layers.62.self_attn.q_a_proj": {
2189
+ "bits": 8
2190
+ },
2191
+ "model.layers.62.self_attn.q_b_proj": {
2192
+ "bits": 8
2193
+ },
2194
+ "model.layers.63.mlp.shared_experts.down_proj": {
2195
+ "bits": 8
2196
+ },
2197
+ "model.layers.63.mlp.shared_experts.gate_proj": {
2198
+ "bits": 8
2199
+ },
2200
+ "model.layers.63.mlp.shared_experts.up_proj": {
2201
+ "bits": 8
2202
+ },
2203
+ "model.layers.63.self_attn.indexer.weights_proj": {
2204
+ "bits": 16,
2205
+ "data_type": "float"
2206
+ },
2207
+ "model.layers.63.self_attn.indexer.wk": {
2208
+ "bits": 8
2209
+ },
2210
+ "model.layers.63.self_attn.indexer.wq_b": {
2211
+ "bits": 8
2212
+ },
2213
+ "model.layers.63.self_attn.kv_a_proj_with_mqa": {
2214
+ "bits": 8
2215
+ },
2216
+ "model.layers.63.self_attn.kv_b_proj": {
2217
+ "bits": 8
2218
+ },
2219
+ "model.layers.63.self_attn.o_proj": {
2220
+ "bits": 8
2221
+ },
2222
+ "model.layers.63.self_attn.q_a_proj": {
2223
+ "bits": 8
2224
+ },
2225
+ "model.layers.63.self_attn.q_b_proj": {
2226
+ "bits": 8
2227
+ },
2228
+ "model.layers.64.mlp.shared_experts.down_proj": {
2229
+ "bits": 8
2230
+ },
2231
+ "model.layers.64.mlp.shared_experts.gate_proj": {
2232
+ "bits": 8
2233
+ },
2234
+ "model.layers.64.mlp.shared_experts.up_proj": {
2235
+ "bits": 8
2236
+ },
2237
+ "model.layers.64.self_attn.indexer.weights_proj": {
2238
+ "bits": 16,
2239
+ "data_type": "float"
2240
+ },
2241
+ "model.layers.64.self_attn.indexer.wk": {
2242
+ "bits": 8
2243
+ },
2244
+ "model.layers.64.self_attn.indexer.wq_b": {
2245
+ "bits": 8
2246
+ },
2247
+ "model.layers.64.self_attn.kv_a_proj_with_mqa": {
2248
+ "bits": 8
2249
+ },
2250
+ "model.layers.64.self_attn.kv_b_proj": {
2251
+ "bits": 8
2252
+ },
2253
+ "model.layers.64.self_attn.o_proj": {
2254
+ "bits": 8
2255
+ },
2256
+ "model.layers.64.self_attn.q_a_proj": {
2257
+ "bits": 8
2258
+ },
2259
+ "model.layers.64.self_attn.q_b_proj": {
2260
+ "bits": 8
2261
+ },
2262
+ "model.layers.65.mlp.shared_experts.down_proj": {
2263
+ "bits": 8
2264
+ },
2265
+ "model.layers.65.mlp.shared_experts.gate_proj": {
2266
+ "bits": 8
2267
+ },
2268
+ "model.layers.65.mlp.shared_experts.up_proj": {
2269
+ "bits": 8
2270
+ },
2271
+ "model.layers.65.self_attn.indexer.weights_proj": {
2272
+ "bits": 16,
2273
+ "data_type": "float"
2274
+ },
2275
+ "model.layers.65.self_attn.indexer.wk": {
2276
+ "bits": 8
2277
+ },
2278
+ "model.layers.65.self_attn.indexer.wq_b": {
2279
+ "bits": 8
2280
+ },
2281
+ "model.layers.65.self_attn.kv_a_proj_with_mqa": {
2282
+ "bits": 8
2283
+ },
2284
+ "model.layers.65.self_attn.kv_b_proj": {
2285
+ "bits": 8
2286
+ },
2287
+ "model.layers.65.self_attn.o_proj": {
2288
+ "bits": 8
2289
+ },
2290
+ "model.layers.65.self_attn.q_a_proj": {
2291
+ "bits": 8
2292
+ },
2293
+ "model.layers.65.self_attn.q_b_proj": {
2294
+ "bits": 8
2295
+ },
2296
+ "model.layers.66.mlp.shared_experts.down_proj": {
2297
+ "bits": 8
2298
+ },
2299
+ "model.layers.66.mlp.shared_experts.gate_proj": {
2300
+ "bits": 8
2301
+ },
2302
+ "model.layers.66.mlp.shared_experts.up_proj": {
2303
+ "bits": 8
2304
+ },
2305
+ "model.layers.66.self_attn.indexer.weights_proj": {
2306
+ "bits": 16,
2307
+ "data_type": "float"
2308
+ },
2309
+ "model.layers.66.self_attn.indexer.wk": {
2310
+ "bits": 8
2311
+ },
2312
+ "model.layers.66.self_attn.indexer.wq_b": {
2313
+ "bits": 8
2314
+ },
2315
+ "model.layers.66.self_attn.kv_a_proj_with_mqa": {
2316
+ "bits": 8
2317
+ },
2318
+ "model.layers.66.self_attn.kv_b_proj": {
2319
+ "bits": 8
2320
+ },
2321
+ "model.layers.66.self_attn.o_proj": {
2322
+ "bits": 8
2323
+ },
2324
+ "model.layers.66.self_attn.q_a_proj": {
2325
+ "bits": 8
2326
+ },
2327
+ "model.layers.66.self_attn.q_b_proj": {
2328
+ "bits": 8
2329
+ },
2330
+ "model.layers.67.mlp.shared_experts.down_proj": {
2331
+ "bits": 8
2332
+ },
2333
+ "model.layers.67.mlp.shared_experts.gate_proj": {
2334
+ "bits": 8
2335
+ },
2336
+ "model.layers.67.mlp.shared_experts.up_proj": {
2337
+ "bits": 8
2338
+ },
2339
+ "model.layers.67.self_attn.indexer.weights_proj": {
2340
+ "bits": 16,
2341
+ "data_type": "float"
2342
+ },
2343
+ "model.layers.67.self_attn.indexer.wk": {
2344
+ "bits": 8
2345
+ },
2346
+ "model.layers.67.self_attn.indexer.wq_b": {
2347
+ "bits": 8
2348
+ },
2349
+ "model.layers.67.self_attn.kv_a_proj_with_mqa": {
2350
+ "bits": 8
2351
+ },
2352
+ "model.layers.67.self_attn.kv_b_proj": {
2353
+ "bits": 8
2354
+ },
2355
+ "model.layers.67.self_attn.o_proj": {
2356
+ "bits": 8
2357
+ },
2358
+ "model.layers.67.self_attn.q_a_proj": {
2359
+ "bits": 8
2360
+ },
2361
+ "model.layers.67.self_attn.q_b_proj": {
2362
+ "bits": 8
2363
+ },
2364
+ "model.layers.68.mlp.shared_experts.down_proj": {
2365
+ "bits": 8
2366
+ },
2367
+ "model.layers.68.mlp.shared_experts.gate_proj": {
2368
+ "bits": 8
2369
+ },
2370
+ "model.layers.68.mlp.shared_experts.up_proj": {
2371
+ "bits": 8
2372
+ },
2373
+ "model.layers.68.self_attn.indexer.weights_proj": {
2374
+ "bits": 16,
2375
+ "data_type": "float"
2376
+ },
2377
+ "model.layers.68.self_attn.indexer.wk": {
2378
+ "bits": 8
2379
+ },
2380
+ "model.layers.68.self_attn.indexer.wq_b": {
2381
+ "bits": 8
2382
+ },
2383
+ "model.layers.68.self_attn.kv_a_proj_with_mqa": {
2384
+ "bits": 8
2385
+ },
2386
+ "model.layers.68.self_attn.kv_b_proj": {
2387
+ "bits": 8
2388
+ },
2389
+ "model.layers.68.self_attn.o_proj": {
2390
+ "bits": 8
2391
+ },
2392
+ "model.layers.68.self_attn.q_a_proj": {
2393
+ "bits": 8
2394
+ },
2395
+ "model.layers.68.self_attn.q_b_proj": {
2396
+ "bits": 8
2397
+ },
2398
+ "model.layers.69.mlp.shared_experts.down_proj": {
2399
+ "bits": 8
2400
+ },
2401
+ "model.layers.69.mlp.shared_experts.gate_proj": {
2402
+ "bits": 8
2403
+ },
2404
+ "model.layers.69.mlp.shared_experts.up_proj": {
2405
+ "bits": 8
2406
+ },
2407
+ "model.layers.69.self_attn.indexer.weights_proj": {
2408
+ "bits": 16,
2409
+ "data_type": "float"
2410
+ },
2411
+ "model.layers.69.self_attn.indexer.wk": {
2412
+ "bits": 8
2413
+ },
2414
+ "model.layers.69.self_attn.indexer.wq_b": {
2415
+ "bits": 8
2416
+ },
2417
+ "model.layers.69.self_attn.kv_a_proj_with_mqa": {
2418
+ "bits": 8
2419
+ },
2420
+ "model.layers.69.self_attn.kv_b_proj": {
2421
+ "bits": 8
2422
+ },
2423
+ "model.layers.69.self_attn.o_proj": {
2424
+ "bits": 8
2425
+ },
2426
+ "model.layers.69.self_attn.q_a_proj": {
2427
+ "bits": 8
2428
+ },
2429
+ "model.layers.69.self_attn.q_b_proj": {
2430
+ "bits": 8
2431
+ },
2432
+ "model.layers.7.mlp.shared_experts.down_proj": {
2433
+ "bits": 8
2434
+ },
2435
+ "model.layers.7.mlp.shared_experts.gate_proj": {
2436
+ "bits": 8
2437
+ },
2438
+ "model.layers.7.mlp.shared_experts.up_proj": {
2439
+ "bits": 8
2440
+ },
2441
+ "model.layers.7.self_attn.indexer.weights_proj": {
2442
+ "bits": 16,
2443
+ "data_type": "float"
2444
+ },
2445
+ "model.layers.7.self_attn.indexer.wk": {
2446
+ "bits": 8
2447
+ },
2448
+ "model.layers.7.self_attn.indexer.wq_b": {
2449
+ "bits": 8
2450
+ },
2451
+ "model.layers.7.self_attn.kv_a_proj_with_mqa": {
2452
+ "bits": 8
2453
+ },
2454
+ "model.layers.7.self_attn.kv_b_proj": {
2455
+ "bits": 8
2456
+ },
2457
+ "model.layers.7.self_attn.o_proj": {
2458
+ "bits": 8
2459
+ },
2460
+ "model.layers.7.self_attn.q_a_proj": {
2461
+ "bits": 8
2462
+ },
2463
+ "model.layers.7.self_attn.q_b_proj": {
2464
+ "bits": 8
2465
+ },
2466
+ "model.layers.70.mlp.shared_experts.down_proj": {
2467
+ "bits": 8
2468
+ },
2469
+ "model.layers.70.mlp.shared_experts.gate_proj": {
2470
+ "bits": 8
2471
+ },
2472
+ "model.layers.70.mlp.shared_experts.up_proj": {
2473
+ "bits": 8
2474
+ },
2475
+ "model.layers.70.self_attn.indexer.weights_proj": {
2476
+ "bits": 16,
2477
+ "data_type": "float"
2478
+ },
2479
+ "model.layers.70.self_attn.indexer.wk": {
2480
+ "bits": 8
2481
+ },
2482
+ "model.layers.70.self_attn.indexer.wq_b": {
2483
+ "bits": 8
2484
+ },
2485
+ "model.layers.70.self_attn.kv_a_proj_with_mqa": {
2486
+ "bits": 8
2487
+ },
2488
+ "model.layers.70.self_attn.kv_b_proj": {
2489
+ "bits": 8
2490
+ },
2491
+ "model.layers.70.self_attn.o_proj": {
2492
+ "bits": 8
2493
+ },
2494
+ "model.layers.70.self_attn.q_a_proj": {
2495
+ "bits": 8
2496
+ },
2497
+ "model.layers.70.self_attn.q_b_proj": {
2498
+ "bits": 8
2499
+ },
2500
+ "model.layers.71.mlp.shared_experts.down_proj": {
2501
+ "bits": 8
2502
+ },
2503
+ "model.layers.71.mlp.shared_experts.gate_proj": {
2504
+ "bits": 8
2505
+ },
2506
+ "model.layers.71.mlp.shared_experts.up_proj": {
2507
+ "bits": 8
2508
+ },
2509
+ "model.layers.71.self_attn.indexer.weights_proj": {
2510
+ "bits": 16,
2511
+ "data_type": "float"
2512
+ },
2513
+ "model.layers.71.self_attn.indexer.wk": {
2514
+ "bits": 8
2515
+ },
2516
+ "model.layers.71.self_attn.indexer.wq_b": {
2517
+ "bits": 8
2518
+ },
2519
+ "model.layers.71.self_attn.kv_a_proj_with_mqa": {
2520
+ "bits": 8
2521
+ },
2522
+ "model.layers.71.self_attn.kv_b_proj": {
2523
+ "bits": 8
2524
+ },
2525
+ "model.layers.71.self_attn.o_proj": {
2526
+ "bits": 8
2527
+ },
2528
+ "model.layers.71.self_attn.q_a_proj": {
2529
+ "bits": 8
2530
+ },
2531
+ "model.layers.71.self_attn.q_b_proj": {
2532
+ "bits": 8
2533
+ },
2534
+ "model.layers.72.mlp.shared_experts.down_proj": {
2535
+ "bits": 8
2536
+ },
2537
+ "model.layers.72.mlp.shared_experts.gate_proj": {
2538
+ "bits": 8
2539
+ },
2540
+ "model.layers.72.mlp.shared_experts.up_proj": {
2541
+ "bits": 8
2542
+ },
2543
+ "model.layers.72.self_attn.indexer.weights_proj": {
2544
+ "bits": 16,
2545
+ "data_type": "float"
2546
+ },
2547
+ "model.layers.72.self_attn.indexer.wk": {
2548
+ "bits": 8
2549
+ },
2550
+ "model.layers.72.self_attn.indexer.wq_b": {
2551
+ "bits": 8
2552
+ },
2553
+ "model.layers.72.self_attn.kv_a_proj_with_mqa": {
2554
+ "bits": 8
2555
+ },
2556
+ "model.layers.72.self_attn.kv_b_proj": {
2557
+ "bits": 8
2558
+ },
2559
+ "model.layers.72.self_attn.o_proj": {
2560
+ "bits": 8
2561
+ },
2562
+ "model.layers.72.self_attn.q_a_proj": {
2563
+ "bits": 8
2564
+ },
2565
+ "model.layers.72.self_attn.q_b_proj": {
2566
+ "bits": 8
2567
+ },
2568
+ "model.layers.73.mlp.shared_experts.down_proj": {
2569
+ "bits": 8
2570
+ },
2571
+ "model.layers.73.mlp.shared_experts.gate_proj": {
2572
+ "bits": 8
2573
+ },
2574
+ "model.layers.73.mlp.shared_experts.up_proj": {
2575
+ "bits": 8
2576
+ },
2577
+ "model.layers.73.self_attn.indexer.weights_proj": {
2578
+ "bits": 16,
2579
+ "data_type": "float"
2580
+ },
2581
+ "model.layers.73.self_attn.indexer.wk": {
2582
+ "bits": 8
2583
+ },
2584
+ "model.layers.73.self_attn.indexer.wq_b": {
2585
+ "bits": 8
2586
+ },
2587
+ "model.layers.73.self_attn.kv_a_proj_with_mqa": {
2588
+ "bits": 8
2589
+ },
2590
+ "model.layers.73.self_attn.kv_b_proj": {
2591
+ "bits": 8
2592
+ },
2593
+ "model.layers.73.self_attn.o_proj": {
2594
+ "bits": 8
2595
+ },
2596
+ "model.layers.73.self_attn.q_a_proj": {
2597
+ "bits": 8
2598
+ },
2599
+ "model.layers.73.self_attn.q_b_proj": {
2600
+ "bits": 8
2601
+ },
2602
+ "model.layers.74.mlp.shared_experts.down_proj": {
2603
+ "bits": 8
2604
+ },
2605
+ "model.layers.74.mlp.shared_experts.gate_proj": {
2606
+ "bits": 8
2607
+ },
2608
+ "model.layers.74.mlp.shared_experts.up_proj": {
2609
+ "bits": 8
2610
+ },
2611
+ "model.layers.74.self_attn.indexer.weights_proj": {
2612
+ "bits": 16,
2613
+ "data_type": "float"
2614
+ },
2615
+ "model.layers.74.self_attn.indexer.wk": {
2616
+ "bits": 8
2617
+ },
2618
+ "model.layers.74.self_attn.indexer.wq_b": {
2619
+ "bits": 8
2620
+ },
2621
+ "model.layers.74.self_attn.kv_a_proj_with_mqa": {
2622
+ "bits": 8
2623
+ },
2624
+ "model.layers.74.self_attn.kv_b_proj": {
2625
+ "bits": 8
2626
+ },
2627
+ "model.layers.74.self_attn.o_proj": {
2628
+ "bits": 8
2629
+ },
2630
+ "model.layers.74.self_attn.q_a_proj": {
2631
+ "bits": 8
2632
+ },
2633
+ "model.layers.74.self_attn.q_b_proj": {
2634
+ "bits": 8
2635
+ },
2636
+ "model.layers.75.mlp.shared_experts.down_proj": {
2637
+ "bits": 8
2638
+ },
2639
+ "model.layers.75.mlp.shared_experts.gate_proj": {
2640
+ "bits": 8
2641
+ },
2642
+ "model.layers.75.mlp.shared_experts.up_proj": {
2643
+ "bits": 8
2644
+ },
2645
+ "model.layers.75.self_attn.indexer.weights_proj": {
2646
+ "bits": 16,
2647
+ "data_type": "float"
2648
+ },
2649
+ "model.layers.75.self_attn.indexer.wk": {
2650
+ "bits": 8
2651
+ },
2652
+ "model.layers.75.self_attn.indexer.wq_b": {
2653
+ "bits": 8
2654
+ },
2655
+ "model.layers.75.self_attn.kv_a_proj_with_mqa": {
2656
+ "bits": 8
2657
+ },
2658
+ "model.layers.75.self_attn.kv_b_proj": {
2659
+ "bits": 8
2660
+ },
2661
+ "model.layers.75.self_attn.o_proj": {
2662
+ "bits": 8
2663
+ },
2664
+ "model.layers.75.self_attn.q_a_proj": {
2665
+ "bits": 8
2666
+ },
2667
+ "model.layers.75.self_attn.q_b_proj": {
2668
+ "bits": 8
2669
+ },
2670
+ "model.layers.76.mlp.shared_experts.down_proj": {
2671
+ "bits": 8
2672
+ },
2673
+ "model.layers.76.mlp.shared_experts.gate_proj": {
2674
+ "bits": 8
2675
+ },
2676
+ "model.layers.76.mlp.shared_experts.up_proj": {
2677
+ "bits": 8
2678
+ },
2679
+ "model.layers.76.self_attn.indexer.weights_proj": {
2680
+ "bits": 16,
2681
+ "data_type": "float"
2682
+ },
2683
+ "model.layers.76.self_attn.indexer.wk": {
2684
+ "bits": 8
2685
+ },
2686
+ "model.layers.76.self_attn.indexer.wq_b": {
2687
+ "bits": 8
2688
+ },
2689
+ "model.layers.76.self_attn.kv_a_proj_with_mqa": {
2690
+ "bits": 8
2691
+ },
2692
+ "model.layers.76.self_attn.kv_b_proj": {
2693
+ "bits": 8
2694
+ },
2695
+ "model.layers.76.self_attn.o_proj": {
2696
+ "bits": 8
2697
+ },
2698
+ "model.layers.76.self_attn.q_a_proj": {
2699
+ "bits": 8
2700
+ },
2701
+ "model.layers.76.self_attn.q_b_proj": {
2702
+ "bits": 8
2703
+ },
2704
+ "model.layers.77.mlp.shared_experts.down_proj": {
2705
+ "bits": 8
2706
+ },
2707
+ "model.layers.77.mlp.shared_experts.gate_proj": {
2708
+ "bits": 8
2709
+ },
2710
+ "model.layers.77.mlp.shared_experts.up_proj": {
2711
+ "bits": 8
2712
+ },
2713
+ "model.layers.77.self_attn.indexer.weights_proj": {
2714
+ "bits": 16,
2715
+ "data_type": "float"
2716
+ },
2717
+ "model.layers.77.self_attn.indexer.wk": {
2718
+ "bits": 8
2719
+ },
2720
+ "model.layers.77.self_attn.indexer.wq_b": {
2721
+ "bits": 8
2722
+ },
2723
+ "model.layers.77.self_attn.kv_a_proj_with_mqa": {
2724
+ "bits": 8
2725
+ },
2726
+ "model.layers.77.self_attn.kv_b_proj": {
2727
+ "bits": 8
2728
+ },
2729
+ "model.layers.77.self_attn.o_proj": {
2730
+ "bits": 8
2731
+ },
2732
+ "model.layers.77.self_attn.q_a_proj": {
2733
+ "bits": 8
2734
+ },
2735
+ "model.layers.77.self_attn.q_b_proj": {
2736
+ "bits": 8
2737
+ },
2738
+ "model.layers.8.mlp.shared_experts.down_proj": {
2739
+ "bits": 8
2740
+ },
2741
+ "model.layers.8.mlp.shared_experts.gate_proj": {
2742
+ "bits": 8
2743
+ },
2744
+ "model.layers.8.mlp.shared_experts.up_proj": {
2745
+ "bits": 8
2746
+ },
2747
+ "model.layers.8.self_attn.indexer.weights_proj": {
2748
+ "bits": 16,
2749
+ "data_type": "float"
2750
+ },
2751
+ "model.layers.8.self_attn.indexer.wk": {
2752
+ "bits": 8
2753
+ },
2754
+ "model.layers.8.self_attn.indexer.wq_b": {
2755
+ "bits": 8
2756
+ },
2757
+ "model.layers.8.self_attn.kv_a_proj_with_mqa": {
2758
+ "bits": 8
2759
+ },
2760
+ "model.layers.8.self_attn.kv_b_proj": {
2761
+ "bits": 8
2762
+ },
2763
+ "model.layers.8.self_attn.o_proj": {
2764
+ "bits": 8
2765
+ },
2766
+ "model.layers.8.self_attn.q_a_proj": {
2767
+ "bits": 8
2768
+ },
2769
+ "model.layers.8.self_attn.q_b_proj": {
2770
+ "bits": 8
2771
+ },
2772
+ "model.layers.9.mlp.shared_experts.down_proj": {
2773
+ "bits": 8
2774
+ },
2775
+ "model.layers.9.mlp.shared_experts.gate_proj": {
2776
+ "bits": 8
2777
+ },
2778
+ "model.layers.9.mlp.shared_experts.up_proj": {
2779
+ "bits": 8
2780
+ },
2781
+ "model.layers.9.self_attn.indexer.weights_proj": {
2782
+ "bits": 16,
2783
+ "data_type": "float"
2784
+ },
2785
+ "model.layers.9.self_attn.indexer.wk": {
2786
+ "bits": 8
2787
+ },
2788
+ "model.layers.9.self_attn.indexer.wq_b": {
2789
+ "bits": 8
2790
+ },
2791
+ "model.layers.9.self_attn.kv_a_proj_with_mqa": {
2792
+ "bits": 8
2793
+ },
2794
+ "model.layers.9.self_attn.kv_b_proj": {
2795
+ "bits": 8
2796
+ },
2797
+ "model.layers.9.self_attn.o_proj": {
2798
+ "bits": 8
2799
+ },
2800
+ "model.layers.9.self_attn.q_a_proj": {
2801
+ "bits": 8
2802
+ },
2803
+ "model.layers.9.self_attn.q_b_proj": {
2804
+ "bits": 8
2805
+ },
2806
+ "model.layers.78.mlp.gate": {
2807
+ "bits": 16,
2808
+ "data_type": "fp"
2809
+ },
2810
+ "model.layers.78.self_attn.indexer.weights_proj": {
2811
+ "bits": 16,
2812
+ "data_type": "fp"
2813
+ }
2814
+ },
2815
+ "group_size": 128,
2816
+ "iters": 0,
2817
+ "packing_format": "auto_round:auto_gptq",
2818
+ "quant_method": "auto-round",
2819
+ "sym": true,
2820
+ "block_name_to_quantize": [
2821
+ "model.layers"
2822
+ ]
2823
+ },
2824
+ "rms_norm_eps": 1e-05,
2825
+ "rope_interleave": true,
2826
+ "rope_parameters": {
2827
+ "rope_theta": 1000000,
2828
+ "rope_type": "default"
2829
+ },
2830
+ "routed_scaling_factor": 2.5,
2831
+ "scoring_func": "sigmoid",
2832
+ "tie_word_embeddings": false,
2833
+ "topk_group": 1,
2834
+ "topk_method": "noaux_tc",
2835
+ "transformers_version": "5.4.0",
2836
+ "use_cache": true,
2837
+ "v_head_dim": 256,
2838
+ "vocab_size": 154880
2839
+ }
generation_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "do_sample": true,
4
+ "eos_token_id": [
5
+ 154820,
6
+ 154827,
7
+ 154829
8
+ ],
9
+ "pad_token_id": 154820,
10
+ "temperature": 1.0,
11
+ "top_p": 0.95,
12
+ "transformers_version": "5.4.0"
13
+ }
model-00001-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3abae9a7f485f43691fcf7f6d6615340ebd6cae24a1390166f24a33ad31ebe84
3
+ size 5368688176
model-00002-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:811fb060d00f21257226be35ca2658d5b9f09f4f2750f4a582acbec30cd429d7
3
+ size 5365562632
model-00003-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2211dfb9d98139969d37f2fa43d67a043ec15406038b4a36a62b550ce8abd81
3
+ size 5365562624
model-00004-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9803461fc68de67edb1b1daec21061984362b4a8d68ea49416170169ea19ac47
3
+ size 5365562632
model-00005-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2cbf72e8e4ef69c1f47645ff4362dd6982485710945c5b481398b644d7acfcae
3
+ size 5365562632
model-00006-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9e8088d7e1f4579eabfc8f6251c5cee53c4294b6c889cd7d86f9c5672b6db68
3
+ size 5365562624
model-00007-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ec63625917b57777d6d9c02af3c115d800c2188d1edb13ebb34025174160c96b
3
+ size 5365562632
model-00008-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7dd89b34bdb47de415d1040f630b5a05e9a9e18d350eda502767785c27024164
3
+ size 5365564568
model-00009-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fc55f84d8e701fe0496a0f23c34d0a9bf8fe342352bd374e7dc1e91e401f167e
3
+ size 5365565040
model-00010-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:812295dc6cd2d21b422ba7c15d03e7a398b2c34d50d82404636e002bef8f2a95
3
+ size 5365565048
model-00011-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:486213678345d3a27fd0a4e6bac1972bfe8d380b430bfc72f6015319fac4e971
3
+ size 5365565048
model-00012-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7f68011f12624ec7294c273904ddd4df98a1002c81f7599e7633689beb9c880a
3
+ size 5365565040
model-00013-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:48adac517365a2ca6a8c1f5a7524ecec164ad752ce4337f91d1fe19127b18195
3
+ size 5365565048
model-00014-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:72b448add7877c93da80a139c8043f30942ea72d7c2b8be08ee04af7614c6f6c
3
+ size 5365565048
model-00015-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ccdabf94269f794c8706a245b77e0e9eff46a0120b9d663cf78d4d78e97b6b56
3
+ size 5368515040
model-00016-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6cc9c53205684d0312e13273c3ef3cfa83fac45305bf2ce4d702a55c9b712c2
3
+ size 5333003624
model-00017-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b8277bf4b5b71152264e8cf3c36937954f983bb97f7c15422dd988971a89e697
3
+ size 5367233880
model-00018-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2aa6322300f1ed1c1ad647a49d0e866d8d9149ce2017787d018468aad0fb0924
3
+ size 5365564920
model-00019-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:844d63a3542fe1d0fc554f34391891e7e8aeeb3b607e0df58970aaf29caa5885
3
+ size 5365564952
model-00020-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:28ac57596a415490b6395949b04748a10a7c34f8550a5eca7110cee9837cb8e6
3
+ size 5365564968
model-00021-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b5e9322c91e504018edc1eaf98bc8f12e89ff65d8185e4e7d8b517e316f276d5
3
+ size 5365564976
model-00022-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ff9abc96d3b07f604dff23668e5e2b25b613c957dc75173a780e61057284d9e9
3
+ size 5365564976
model-00023-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8898c99133a028a488eacfff87f363caaf918576a06d9f29eb95310b3e21384d
3
+ size 5365564968
model-00024-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:dd6c3d794cac63b8d20058438428ff10e8c99e91e768c72dbdf9b846c5d3d473
3
+ size 5365564976
model-00025-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:51e7cfc3c52229c749822117a276ebb2760fb9d83e9e1c6ebce3e262cc9db7ed
3
+ size 5365564976
model-00026-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e6ce74b934bce565399fab2918027ddd184cabeab1eff1d663f7e3ead36052c4
3
+ size 5365564968
model-00027-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b9230049da8accd29f58b7e3bed2abae63f44f0cf02a5bda56588b57485a956
3
+ size 5365564976
model-00028-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b1752884eb7c6b1dd6894715ce767fe86c267bc075ce918a349e005c9932841
3
+ size 5365564976
model-00029-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6c0d17423b427e4aa8c73c17d97d26d3414858833971dbf0071938efaf794eeb
3
+ size 5365564968
model-00030-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:07211a77e3332004912734b3f2cd6361ebff2dc5c759ea44d3e8e3a23bd1816a
3
+ size 5365564976
model-00031-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:80deb888b7718e3a2bc8898f35633832084e1ab9fa957dc5b51bb719354c5524
3
+ size 5365564976
model-00032-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7114e5eb37ff04de14ecac90117b54d946d6e50e33f7cef359e9e3b49defa39
3
+ size 5365564968
model-00033-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4483ba090cdd3a230c983fc1f377ac7dbaee1d5ed30721bffd3dd9ed0a3fd59e
3
+ size 5365564992
model-00034-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf6003585a116927d18522b18f89b417d38e8fe465f60c890fe37753b5d5d846
3
+ size 5365565032
model-00035-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42893ee9b9046f1777fce11674be8ac6b5214e507d7ba9295c8424e3fec53383
3
+ size 5365565024
model-00036-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aa3d30b4f3111e71cba4e9fdeaf21a1b82fb0e0c60694eef40b2cb2da6d787fd
3
+ size 5365565032
model-00037-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:079b1d0dc9bb5e09e0d9c9aaa1c851d6e54fa78f726bda586afe0c8e4a2ce86e
3
+ size 5365565032
model-00038-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:53aa1a901f7668c805b32290fb6b12b84cb130a9feac4f6dcc44c3de2aecae20
3
+ size 5365565024
model-00039-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7ce6782494a0e95cbf884c3e46ba34628f7e59933597545e1f2ac2fc1c17177d
3
+ size 5365565032
model-00040-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:630a428b4ce8714f05bd09bad190f98a983393fbde07410a0580426df6a841d2
3
+ size 5365565032
model-00041-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9fa25b73a8779ecdfd595f3ff850ac7aa0c6d15477a8772a62b4aa3d65c61e44
3
+ size 5365565024
model-00042-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7c681f82a0f42547d01191ddf7478d305725c3cf9689df8016b50695c67439f
3
+ size 5365565032
model-00043-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0a6c3284e9c31940f4ae1ff4611862a35cb357594c1a34a6ef7fb4105909189e
3
+ size 5365565032
model-00044-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8800dc8738568b5f7eb2b8f56525381cbd0ac7daaaf93374c1d52a2db5ef20e8
3
+ size 5365565024
model-00045-of-00075.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc07e23395c3d3d30fe8352b4687676420c778d0e2ab6e6f476657d3293e08a7
3
+ size 5365565032