This view is limited to 50 files because it contains too many changes. See the raw diff here.
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. chat_template.jinja +112 -0
  3. config.json +712 -0
  4. configuration_deepseek.py +214 -0
  5. configuration_kimi_k25.py +123 -0
  6. generation_config.json +8 -0
  7. kimi_k25_vision_processing.py +251 -0
  8. media_utils.py +368 -0
  9. model-00001-of-00119.safetensors +3 -0
  10. model-00002-of-00119.safetensors +3 -0
  11. model-00003-of-00119.safetensors +3 -0
  12. model-00004-of-00119.safetensors +3 -0
  13. model-00005-of-00119.safetensors +3 -0
  14. model-00006-of-00119.safetensors +3 -0
  15. model-00007-of-00119.safetensors +3 -0
  16. model-00008-of-00119.safetensors +3 -0
  17. model-00009-of-00119.safetensors +3 -0
  18. model-00010-of-00119.safetensors +3 -0
  19. model-00011-of-00119.safetensors +3 -0
  20. model-00012-of-00119.safetensors +3 -0
  21. model-00013-of-00119.safetensors +3 -0
  22. model-00014-of-00119.safetensors +3 -0
  23. model-00015-of-00119.safetensors +3 -0
  24. model-00016-of-00119.safetensors +3 -0
  25. model-00017-of-00119.safetensors +3 -0
  26. model-00018-of-00119.safetensors +3 -0
  27. model-00019-of-00119.safetensors +3 -0
  28. model-00020-of-00119.safetensors +3 -0
  29. model-00021-of-00119.safetensors +3 -0
  30. model-00022-of-00119.safetensors +3 -0
  31. model-00023-of-00119.safetensors +3 -0
  32. model-00024-of-00119.safetensors +3 -0
  33. model-00025-of-00119.safetensors +3 -0
  34. model-00026-of-00119.safetensors +3 -0
  35. model-00027-of-00119.safetensors +3 -0
  36. model-00028-of-00119.safetensors +3 -0
  37. model-00029-of-00119.safetensors +3 -0
  38. model-00030-of-00119.safetensors +3 -0
  39. model-00031-of-00119.safetensors +3 -0
  40. model-00032-of-00119.safetensors +3 -0
  41. model-00033-of-00119.safetensors +3 -0
  42. model-00034-of-00119.safetensors +3 -0
  43. model-00035-of-00119.safetensors +3 -0
  44. model-00036-of-00119.safetensors +3 -0
  45. model-00037-of-00119.safetensors +3 -0
  46. model-00038-of-00119.safetensors +3 -0
  47. model-00039-of-00119.safetensors +3 -0
  48. model-00040-of-00119.safetensors +3 -0
  49. model-00041-of-00119.safetensors +3 -0
  50. model-00042-of-00119.safetensors +3 -0
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
chat_template.jinja ADDED
@@ -0,0 +1,112 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- macro render_content(msg) -%}
2
+ {%- set c = msg.get('content') -%}
3
+ {%- if c is string -%}
4
+ {{ c }}
5
+ {%- elif c is not none -%}
6
+ {% for content in c -%}
7
+ {% if content['type'] == 'image' or content['type'] == 'image_url' -%}
8
+ <|media_begin|>image<|media_content|><|media_pad|><|media_end|>
9
+ {% elif content['type'] == 'video' or content['type']== 'video_url'-%}
10
+ <|kimi_k25_video_placeholder|>
11
+ {% else -%}
12
+ {{ content['text'] }}
13
+ {%- endif -%}
14
+ {%- endfor -%}
15
+ {%- endif -%}
16
+ {%- endmacro -%}
17
+
18
+ {% macro set_roles(message) -%}
19
+ {%- set role_name = message.get('name') or message['role'] -%}
20
+ {%- if message['role'] == 'user' -%}
21
+ <|im_user|>{{role_name}}<|im_middle|>
22
+ {%- elif message['role'] == 'assistant' -%}
23
+ <|im_assistant|>{{role_name}}<|im_middle|>
24
+ {%- else -%}
25
+ <|im_system|>{{role_name}}<|im_middle|>
26
+ {%- endif -%}
27
+ {%- endmacro -%}
28
+
29
+
30
+ {%- macro render_toolcalls(message) -%}
31
+ <|tool_calls_section_begin|>
32
+ {%- for tool_call in message['tool_calls'] -%}
33
+ {%- set formatted_id = tool_call['id'] -%}
34
+ <|tool_call_begin|>{{ formatted_id }}<|tool_call_argument_begin|>{% if tool_call['function']['arguments'] is string %}{{ tool_call['function']['arguments'] }}{% else %}{{ tool_call['function']['arguments'] | tojson }}{% endif %}<|tool_call_end|>
35
+ {%- endfor -%}
36
+ <|tool_calls_section_end|>
37
+ {%- endmacro -%}
38
+
39
+
40
+ {%- set preserve_thinking = preserve_thinking | default(false) -%}
41
+ {# Find last non-tool-call assistant message. If preserve_thinking, keep -1 so hist is empty and all msgs use suffix (retain reasoning). #}
42
+ {%- set ns = namespace(last_non_tool_call_assistant_msg=-1) -%}
43
+ {%- if not preserve_thinking -%}
44
+ {%- for idx in range(messages|length-1, -1, -1) -%}
45
+ {%- if messages[idx]['role'] == 'assistant' and not messages[idx].get('tool_calls') -%}
46
+ {%- set ns.last_non_tool_call_assistant_msg = idx -%}
47
+ {%- break -%}
48
+ {%- endif -%}
49
+ {%- endfor -%}
50
+ {%- endif -%}
51
+
52
+ {# split all messages into history & suffix, reasoning_content in suffix should be reserved.#}
53
+ {%- set hist_msgs = messages[:ns.last_non_tool_call_assistant_msg+1] -%}
54
+ {%- set suffix_msgs = messages[ns.last_non_tool_call_assistant_msg+1:] -%}
55
+
56
+ {%- if tools -%}
57
+ {%- if tools_ts_str -%}
58
+ <|im_system|>tool_declare<|im_middle|>{{ tools_ts_str }}<|im_end|>
59
+ {%- else -%}
60
+ <|im_system|>tool_declare<|im_middle|>{{ tools | tojson(separators=(',', ':')) }}<|im_end|>
61
+ {%- endif -%}
62
+ {%- endif -%}
63
+
64
+
65
+ {%- for message in hist_msgs -%}
66
+ {{set_roles(message)}}
67
+ {%- if message['role'] == 'assistant' -%}
68
+ <think></think>{{render_content(message)}}
69
+ {%- if message.get('tool_calls') -%}
70
+ {{render_toolcalls(message)}}
71
+ {%- endif -%}
72
+ {%- elif message['role'] == 'tool' -%}
73
+ {%- set tool_call_id = message.tool_call_id -%}
74
+ ## Return of {{ tool_call_id }}
75
+ {{render_content(message)}}
76
+ {%- elif message['content'] is not none -%}
77
+ {{render_content(message)}}
78
+ {%- endif -%}
79
+ <|im_end|>
80
+ {%- endfor -%}
81
+
82
+ {%- for message in suffix_msgs -%}
83
+ {{set_roles(message)}}
84
+ {%- if message['role'] == 'assistant' -%}
85
+ {%- if thinking is defined and thinking is false and preserve_thinking is false -%}
86
+ <think></think>{{render_content(message)}}
87
+ {%- else -%}
88
+ {%- set rc = message.get('reasoning', message.get('reasoning_content', '')) -%}
89
+ <think>{{rc}}</think>{{render_content(message)}}
90
+ {%- endif -%}
91
+ {%- if message.get('tool_calls') -%}
92
+ {{render_toolcalls(message)}}
93
+ {%- endif -%}
94
+ {%- elif message['role'] == 'tool' -%}
95
+ {%- set tool_call_id = message.tool_call_id -%}
96
+ ## Return of {{ tool_call_id }}
97
+ {{render_content(message)}}
98
+ {%- elif message['content'] is not none -%}
99
+ {{render_content(message)}}
100
+ {%- endif -%}
101
+ <|im_end|>
102
+ {%- endfor -%}
103
+
104
+
105
+ {%- if add_generation_prompt -%}
106
+ <|im_assistant|>assistant<|im_middle|>
107
+ {%- if thinking is defined and thinking is false -%}
108
+ <think></think>
109
+ {%- else -%}
110
+ <think>
111
+ {%- endif -%}
112
+ {%- endif -%}
config.json ADDED
@@ -0,0 +1,712 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "KimiK25ForConditionalGeneration"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_kimi_k25.KimiK25Config",
7
+ "AutoModel": "modeling_kimi_k25.KimiK25ForConditionalGeneration",
8
+ "AutoModelForCausalLM": "modeling_kimi_k25.KimiK25ForConditionalGeneration"
9
+ },
10
+ "bos_token_id": 163584,
11
+ "dtype": "bfloat16",
12
+ "eos_token_id": 163586,
13
+ "ignore_index": -100,
14
+ "media_placeholder_token_id": 163605,
15
+ "model_type": "kimi_k25",
16
+ "pad_token_id": 163839,
17
+ "quantization_config": {
18
+ "algo_config": null,
19
+ "exclude": [
20
+ "vision_tower.patch_embed.proj",
21
+ "vision_tower.encoder.blocks.0.mlp.fc0",
22
+ "vision_tower.encoder.blocks.0.mlp.fc1",
23
+ "vision_tower.encoder.blocks.0.wqkv",
24
+ "vision_tower.encoder.blocks.0.wo",
25
+ "vision_tower.encoder.blocks.1.mlp.fc0",
26
+ "vision_tower.encoder.blocks.1.mlp.fc1",
27
+ "vision_tower.encoder.blocks.1.wqkv",
28
+ "vision_tower.encoder.blocks.1.wo",
29
+ "vision_tower.encoder.blocks.2.mlp.fc0",
30
+ "vision_tower.encoder.blocks.2.mlp.fc1",
31
+ "vision_tower.encoder.blocks.2.wqkv",
32
+ "vision_tower.encoder.blocks.2.wo",
33
+ "vision_tower.encoder.blocks.3.mlp.fc0",
34
+ "vision_tower.encoder.blocks.3.mlp.fc1",
35
+ "vision_tower.encoder.blocks.3.wqkv",
36
+ "vision_tower.encoder.blocks.3.wo",
37
+ "vision_tower.encoder.blocks.4.mlp.fc0",
38
+ "vision_tower.encoder.blocks.4.mlp.fc1",
39
+ "vision_tower.encoder.blocks.4.wqkv",
40
+ "vision_tower.encoder.blocks.4.wo",
41
+ "vision_tower.encoder.blocks.5.mlp.fc0",
42
+ "vision_tower.encoder.blocks.5.mlp.fc1",
43
+ "vision_tower.encoder.blocks.5.wqkv",
44
+ "vision_tower.encoder.blocks.5.wo",
45
+ "vision_tower.encoder.blocks.6.mlp.fc0",
46
+ "vision_tower.encoder.blocks.6.mlp.fc1",
47
+ "vision_tower.encoder.blocks.6.wqkv",
48
+ "vision_tower.encoder.blocks.6.wo",
49
+ "vision_tower.encoder.blocks.7.mlp.fc0",
50
+ "vision_tower.encoder.blocks.7.mlp.fc1",
51
+ "vision_tower.encoder.blocks.7.wqkv",
52
+ "vision_tower.encoder.blocks.7.wo",
53
+ "vision_tower.encoder.blocks.8.mlp.fc0",
54
+ "vision_tower.encoder.blocks.8.mlp.fc1",
55
+ "vision_tower.encoder.blocks.8.wqkv",
56
+ "vision_tower.encoder.blocks.8.wo",
57
+ "vision_tower.encoder.blocks.9.mlp.fc0",
58
+ "vision_tower.encoder.blocks.9.mlp.fc1",
59
+ "vision_tower.encoder.blocks.9.wqkv",
60
+ "vision_tower.encoder.blocks.9.wo",
61
+ "vision_tower.encoder.blocks.10.mlp.fc0",
62
+ "vision_tower.encoder.blocks.10.mlp.fc1",
63
+ "vision_tower.encoder.blocks.10.wqkv",
64
+ "vision_tower.encoder.blocks.10.wo",
65
+ "vision_tower.encoder.blocks.11.mlp.fc0",
66
+ "vision_tower.encoder.blocks.11.mlp.fc1",
67
+ "vision_tower.encoder.blocks.11.wqkv",
68
+ "vision_tower.encoder.blocks.11.wo",
69
+ "vision_tower.encoder.blocks.12.mlp.fc0",
70
+ "vision_tower.encoder.blocks.12.mlp.fc1",
71
+ "vision_tower.encoder.blocks.12.wqkv",
72
+ "vision_tower.encoder.blocks.12.wo",
73
+ "vision_tower.encoder.blocks.13.mlp.fc0",
74
+ "vision_tower.encoder.blocks.13.mlp.fc1",
75
+ "vision_tower.encoder.blocks.13.wqkv",
76
+ "vision_tower.encoder.blocks.13.wo",
77
+ "vision_tower.encoder.blocks.14.mlp.fc0",
78
+ "vision_tower.encoder.blocks.14.mlp.fc1",
79
+ "vision_tower.encoder.blocks.14.wqkv",
80
+ "vision_tower.encoder.blocks.14.wo",
81
+ "vision_tower.encoder.blocks.15.mlp.fc0",
82
+ "vision_tower.encoder.blocks.15.mlp.fc1",
83
+ "vision_tower.encoder.blocks.15.wqkv",
84
+ "vision_tower.encoder.blocks.15.wo",
85
+ "vision_tower.encoder.blocks.16.mlp.fc0",
86
+ "vision_tower.encoder.blocks.16.mlp.fc1",
87
+ "vision_tower.encoder.blocks.16.wqkv",
88
+ "vision_tower.encoder.blocks.16.wo",
89
+ "vision_tower.encoder.blocks.17.mlp.fc0",
90
+ "vision_tower.encoder.blocks.17.mlp.fc1",
91
+ "vision_tower.encoder.blocks.17.wqkv",
92
+ "vision_tower.encoder.blocks.17.wo",
93
+ "vision_tower.encoder.blocks.18.mlp.fc0",
94
+ "vision_tower.encoder.blocks.18.mlp.fc1",
95
+ "vision_tower.encoder.blocks.18.wqkv",
96
+ "vision_tower.encoder.blocks.18.wo",
97
+ "vision_tower.encoder.blocks.19.mlp.fc0",
98
+ "vision_tower.encoder.blocks.19.mlp.fc1",
99
+ "vision_tower.encoder.blocks.19.wqkv",
100
+ "vision_tower.encoder.blocks.19.wo",
101
+ "vision_tower.encoder.blocks.20.mlp.fc0",
102
+ "vision_tower.encoder.blocks.20.mlp.fc1",
103
+ "vision_tower.encoder.blocks.20.wqkv",
104
+ "vision_tower.encoder.blocks.20.wo",
105
+ "vision_tower.encoder.blocks.21.mlp.fc0",
106
+ "vision_tower.encoder.blocks.21.mlp.fc1",
107
+ "vision_tower.encoder.blocks.21.wqkv",
108
+ "vision_tower.encoder.blocks.21.wo",
109
+ "vision_tower.encoder.blocks.22.mlp.fc0",
110
+ "vision_tower.encoder.blocks.22.mlp.fc1",
111
+ "vision_tower.encoder.blocks.22.wqkv",
112
+ "vision_tower.encoder.blocks.22.wo",
113
+ "vision_tower.encoder.blocks.23.mlp.fc0",
114
+ "vision_tower.encoder.blocks.23.mlp.fc1",
115
+ "vision_tower.encoder.blocks.23.wqkv",
116
+ "vision_tower.encoder.blocks.23.wo",
117
+ "vision_tower.encoder.blocks.24.mlp.fc0",
118
+ "vision_tower.encoder.blocks.24.mlp.fc1",
119
+ "vision_tower.encoder.blocks.24.wqkv",
120
+ "vision_tower.encoder.blocks.24.wo",
121
+ "vision_tower.encoder.blocks.25.mlp.fc0",
122
+ "vision_tower.encoder.blocks.25.mlp.fc1",
123
+ "vision_tower.encoder.blocks.25.wqkv",
124
+ "vision_tower.encoder.blocks.25.wo",
125
+ "vision_tower.encoder.blocks.26.mlp.fc0",
126
+ "vision_tower.encoder.blocks.26.mlp.fc1",
127
+ "vision_tower.encoder.blocks.26.wqkv",
128
+ "vision_tower.encoder.blocks.26.wo",
129
+ "mm_projector.proj.0",
130
+ "mm_projector.proj.2",
131
+ "language_model.model.layers.0.self_attn.q_a_proj",
132
+ "language_model.model.layers.0.self_attn.q_b_proj",
133
+ "language_model.model.layers.0.self_attn.kv_a_proj_with_mqa",
134
+ "language_model.model.layers.0.self_attn.kv_b_proj",
135
+ "language_model.model.layers.0.self_attn.o_proj",
136
+ "language_model.model.layers.1.self_attn.q_a_proj",
137
+ "language_model.model.layers.1.self_attn.q_b_proj",
138
+ "language_model.model.layers.1.self_attn.kv_a_proj_with_mqa",
139
+ "language_model.model.layers.1.self_attn.kv_b_proj",
140
+ "language_model.model.layers.1.self_attn.o_proj",
141
+ "language_model.model.layers.2.self_attn.q_a_proj",
142
+ "language_model.model.layers.2.self_attn.q_b_proj",
143
+ "language_model.model.layers.2.self_attn.kv_a_proj_with_mqa",
144
+ "language_model.model.layers.2.self_attn.kv_b_proj",
145
+ "language_model.model.layers.2.self_attn.o_proj",
146
+ "language_model.model.layers.3.self_attn.q_a_proj",
147
+ "language_model.model.layers.3.self_attn.q_b_proj",
148
+ "language_model.model.layers.3.self_attn.kv_a_proj_with_mqa",
149
+ "language_model.model.layers.3.self_attn.kv_b_proj",
150
+ "language_model.model.layers.3.self_attn.o_proj",
151
+ "language_model.model.layers.4.self_attn.q_a_proj",
152
+ "language_model.model.layers.4.self_attn.q_b_proj",
153
+ "language_model.model.layers.4.self_attn.kv_a_proj_with_mqa",
154
+ "language_model.model.layers.4.self_attn.kv_b_proj",
155
+ "language_model.model.layers.4.self_attn.o_proj",
156
+ "language_model.model.layers.5.self_attn.q_a_proj",
157
+ "language_model.model.layers.5.self_attn.q_b_proj",
158
+ "language_model.model.layers.5.self_attn.kv_a_proj_with_mqa",
159
+ "language_model.model.layers.5.self_attn.kv_b_proj",
160
+ "language_model.model.layers.5.self_attn.o_proj",
161
+ "language_model.model.layers.6.self_attn.q_a_proj",
162
+ "language_model.model.layers.6.self_attn.q_b_proj",
163
+ "language_model.model.layers.6.self_attn.kv_a_proj_with_mqa",
164
+ "language_model.model.layers.6.self_attn.kv_b_proj",
165
+ "language_model.model.layers.6.self_attn.o_proj",
166
+ "language_model.model.layers.7.self_attn.q_a_proj",
167
+ "language_model.model.layers.7.self_attn.q_b_proj",
168
+ "language_model.model.layers.7.self_attn.kv_a_proj_with_mqa",
169
+ "language_model.model.layers.7.self_attn.kv_b_proj",
170
+ "language_model.model.layers.7.self_attn.o_proj",
171
+ "language_model.model.layers.8.self_attn.q_a_proj",
172
+ "language_model.model.layers.8.self_attn.q_b_proj",
173
+ "language_model.model.layers.8.self_attn.kv_a_proj_with_mqa",
174
+ "language_model.model.layers.8.self_attn.kv_b_proj",
175
+ "language_model.model.layers.8.self_attn.o_proj",
176
+ "language_model.model.layers.9.self_attn.q_a_proj",
177
+ "language_model.model.layers.9.self_attn.q_b_proj",
178
+ "language_model.model.layers.9.self_attn.kv_a_proj_with_mqa",
179
+ "language_model.model.layers.9.self_attn.kv_b_proj",
180
+ "language_model.model.layers.9.self_attn.o_proj",
181
+ "language_model.model.layers.10.self_attn.q_a_proj",
182
+ "language_model.model.layers.10.self_attn.q_b_proj",
183
+ "language_model.model.layers.10.self_attn.kv_a_proj_with_mqa",
184
+ "language_model.model.layers.10.self_attn.kv_b_proj",
185
+ "language_model.model.layers.10.self_attn.o_proj",
186
+ "language_model.model.layers.11.self_attn.q_a_proj",
187
+ "language_model.model.layers.11.self_attn.q_b_proj",
188
+ "language_model.model.layers.11.self_attn.kv_a_proj_with_mqa",
189
+ "language_model.model.layers.11.self_attn.kv_b_proj",
190
+ "language_model.model.layers.11.self_attn.o_proj",
191
+ "language_model.model.layers.12.self_attn.q_a_proj",
192
+ "language_model.model.layers.12.self_attn.q_b_proj",
193
+ "language_model.model.layers.12.self_attn.kv_a_proj_with_mqa",
194
+ "language_model.model.layers.12.self_attn.kv_b_proj",
195
+ "language_model.model.layers.12.self_attn.o_proj",
196
+ "language_model.model.layers.13.self_attn.q_a_proj",
197
+ "language_model.model.layers.13.self_attn.q_b_proj",
198
+ "language_model.model.layers.13.self_attn.kv_a_proj_with_mqa",
199
+ "language_model.model.layers.13.self_attn.kv_b_proj",
200
+ "language_model.model.layers.13.self_attn.o_proj",
201
+ "language_model.model.layers.14.self_attn.q_a_proj",
202
+ "language_model.model.layers.14.self_attn.q_b_proj",
203
+ "language_model.model.layers.14.self_attn.kv_a_proj_with_mqa",
204
+ "language_model.model.layers.14.self_attn.kv_b_proj",
205
+ "language_model.model.layers.14.self_attn.o_proj",
206
+ "language_model.model.layers.15.self_attn.q_a_proj",
207
+ "language_model.model.layers.15.self_attn.q_b_proj",
208
+ "language_model.model.layers.15.self_attn.kv_a_proj_with_mqa",
209
+ "language_model.model.layers.15.self_attn.kv_b_proj",
210
+ "language_model.model.layers.15.self_attn.o_proj",
211
+ "language_model.model.layers.16.self_attn.q_a_proj",
212
+ "language_model.model.layers.16.self_attn.q_b_proj",
213
+ "language_model.model.layers.16.self_attn.kv_a_proj_with_mqa",
214
+ "language_model.model.layers.16.self_attn.kv_b_proj",
215
+ "language_model.model.layers.16.self_attn.o_proj",
216
+ "language_model.model.layers.17.self_attn.q_a_proj",
217
+ "language_model.model.layers.17.self_attn.q_b_proj",
218
+ "language_model.model.layers.17.self_attn.kv_a_proj_with_mqa",
219
+ "language_model.model.layers.17.self_attn.kv_b_proj",
220
+ "language_model.model.layers.17.self_attn.o_proj",
221
+ "language_model.model.layers.18.self_attn.q_a_proj",
222
+ "language_model.model.layers.18.self_attn.q_b_proj",
223
+ "language_model.model.layers.18.self_attn.kv_a_proj_with_mqa",
224
+ "language_model.model.layers.18.self_attn.kv_b_proj",
225
+ "language_model.model.layers.18.self_attn.o_proj",
226
+ "language_model.model.layers.19.self_attn.q_a_proj",
227
+ "language_model.model.layers.19.self_attn.q_b_proj",
228
+ "language_model.model.layers.19.self_attn.kv_a_proj_with_mqa",
229
+ "language_model.model.layers.19.self_attn.kv_b_proj",
230
+ "language_model.model.layers.19.self_attn.o_proj",
231
+ "language_model.model.layers.20.self_attn.q_a_proj",
232
+ "language_model.model.layers.20.self_attn.q_b_proj",
233
+ "language_model.model.layers.20.self_attn.kv_a_proj_with_mqa",
234
+ "language_model.model.layers.20.self_attn.kv_b_proj",
235
+ "language_model.model.layers.20.self_attn.o_proj",
236
+ "language_model.model.layers.21.self_attn.q_a_proj",
237
+ "language_model.model.layers.21.self_attn.q_b_proj",
238
+ "language_model.model.layers.21.self_attn.kv_a_proj_with_mqa",
239
+ "language_model.model.layers.21.self_attn.kv_b_proj",
240
+ "language_model.model.layers.21.self_attn.o_proj",
241
+ "language_model.model.layers.22.self_attn.q_a_proj",
242
+ "language_model.model.layers.22.self_attn.q_b_proj",
243
+ "language_model.model.layers.22.self_attn.kv_a_proj_with_mqa",
244
+ "language_model.model.layers.22.self_attn.kv_b_proj",
245
+ "language_model.model.layers.22.self_attn.o_proj",
246
+ "language_model.model.layers.23.self_attn.q_a_proj",
247
+ "language_model.model.layers.23.self_attn.q_b_proj",
248
+ "language_model.model.layers.23.self_attn.kv_a_proj_with_mqa",
249
+ "language_model.model.layers.23.self_attn.kv_b_proj",
250
+ "language_model.model.layers.23.self_attn.o_proj",
251
+ "language_model.model.layers.24.self_attn.q_a_proj",
252
+ "language_model.model.layers.24.self_attn.q_b_proj",
253
+ "language_model.model.layers.24.self_attn.kv_a_proj_with_mqa",
254
+ "language_model.model.layers.24.self_attn.kv_b_proj",
255
+ "language_model.model.layers.24.self_attn.o_proj",
256
+ "language_model.model.layers.25.self_attn.q_a_proj",
257
+ "language_model.model.layers.25.self_attn.q_b_proj",
258
+ "language_model.model.layers.25.self_attn.kv_a_proj_with_mqa",
259
+ "language_model.model.layers.25.self_attn.kv_b_proj",
260
+ "language_model.model.layers.25.self_attn.o_proj",
261
+ "language_model.model.layers.26.self_attn.q_a_proj",
262
+ "language_model.model.layers.26.self_attn.q_b_proj",
263
+ "language_model.model.layers.26.self_attn.kv_a_proj_with_mqa",
264
+ "language_model.model.layers.26.self_attn.kv_b_proj",
265
+ "language_model.model.layers.26.self_attn.o_proj",
266
+ "language_model.model.layers.27.self_attn.q_a_proj",
267
+ "language_model.model.layers.27.self_attn.q_b_proj",
268
+ "language_model.model.layers.27.self_attn.kv_a_proj_with_mqa",
269
+ "language_model.model.layers.27.self_attn.kv_b_proj",
270
+ "language_model.model.layers.27.self_attn.o_proj",
271
+ "language_model.model.layers.28.self_attn.q_a_proj",
272
+ "language_model.model.layers.28.self_attn.q_b_proj",
273
+ "language_model.model.layers.28.self_attn.kv_a_proj_with_mqa",
274
+ "language_model.model.layers.28.self_attn.kv_b_proj",
275
+ "language_model.model.layers.28.self_attn.o_proj",
276
+ "language_model.model.layers.29.self_attn.q_a_proj",
277
+ "language_model.model.layers.29.self_attn.q_b_proj",
278
+ "language_model.model.layers.29.self_attn.kv_a_proj_with_mqa",
279
+ "language_model.model.layers.29.self_attn.kv_b_proj",
280
+ "language_model.model.layers.29.self_attn.o_proj",
281
+ "language_model.model.layers.30.self_attn.q_a_proj",
282
+ "language_model.model.layers.30.self_attn.q_b_proj",
283
+ "language_model.model.layers.30.self_attn.kv_a_proj_with_mqa",
284
+ "language_model.model.layers.30.self_attn.kv_b_proj",
285
+ "language_model.model.layers.30.self_attn.o_proj",
286
+ "language_model.model.layers.31.self_attn.q_a_proj",
287
+ "language_model.model.layers.31.self_attn.q_b_proj",
288
+ "language_model.model.layers.31.self_attn.kv_a_proj_with_mqa",
289
+ "language_model.model.layers.31.self_attn.kv_b_proj",
290
+ "language_model.model.layers.31.self_attn.o_proj",
291
+ "language_model.model.layers.32.self_attn.q_a_proj",
292
+ "language_model.model.layers.32.self_attn.q_b_proj",
293
+ "language_model.model.layers.32.self_attn.kv_a_proj_with_mqa",
294
+ "language_model.model.layers.32.self_attn.kv_b_proj",
295
+ "language_model.model.layers.32.self_attn.o_proj",
296
+ "language_model.model.layers.33.self_attn.q_a_proj",
297
+ "language_model.model.layers.33.self_attn.q_b_proj",
298
+ "language_model.model.layers.33.self_attn.kv_a_proj_with_mqa",
299
+ "language_model.model.layers.33.self_attn.kv_b_proj",
300
+ "language_model.model.layers.33.self_attn.o_proj",
301
+ "language_model.model.layers.34.self_attn.q_a_proj",
302
+ "language_model.model.layers.34.self_attn.q_b_proj",
303
+ "language_model.model.layers.34.self_attn.kv_a_proj_with_mqa",
304
+ "language_model.model.layers.34.self_attn.kv_b_proj",
305
+ "language_model.model.layers.34.self_attn.o_proj",
306
+ "language_model.model.layers.35.self_attn.q_a_proj",
307
+ "language_model.model.layers.35.self_attn.q_b_proj",
308
+ "language_model.model.layers.35.self_attn.kv_a_proj_with_mqa",
309
+ "language_model.model.layers.35.self_attn.kv_b_proj",
310
+ "language_model.model.layers.35.self_attn.o_proj",
311
+ "language_model.model.layers.36.self_attn.q_a_proj",
312
+ "language_model.model.layers.36.self_attn.q_b_proj",
313
+ "language_model.model.layers.36.self_attn.kv_a_proj_with_mqa",
314
+ "language_model.model.layers.36.self_attn.kv_b_proj",
315
+ "language_model.model.layers.36.self_attn.o_proj",
316
+ "language_model.model.layers.37.self_attn.q_a_proj",
317
+ "language_model.model.layers.37.self_attn.q_b_proj",
318
+ "language_model.model.layers.37.self_attn.kv_a_proj_with_mqa",
319
+ "language_model.model.layers.37.self_attn.kv_b_proj",
320
+ "language_model.model.layers.37.self_attn.o_proj",
321
+ "language_model.model.layers.38.self_attn.q_a_proj",
322
+ "language_model.model.layers.38.self_attn.q_b_proj",
323
+ "language_model.model.layers.38.self_attn.kv_a_proj_with_mqa",
324
+ "language_model.model.layers.38.self_attn.kv_b_proj",
325
+ "language_model.model.layers.38.self_attn.o_proj",
326
+ "language_model.model.layers.39.self_attn.q_a_proj",
327
+ "language_model.model.layers.39.self_attn.q_b_proj",
328
+ "language_model.model.layers.39.self_attn.kv_a_proj_with_mqa",
329
+ "language_model.model.layers.39.self_attn.kv_b_proj",
330
+ "language_model.model.layers.39.self_attn.o_proj",
331
+ "language_model.model.layers.40.self_attn.q_a_proj",
332
+ "language_model.model.layers.40.self_attn.q_b_proj",
333
+ "language_model.model.layers.40.self_attn.kv_a_proj_with_mqa",
334
+ "language_model.model.layers.40.self_attn.kv_b_proj",
335
+ "language_model.model.layers.40.self_attn.o_proj",
336
+ "language_model.model.layers.41.self_attn.q_a_proj",
337
+ "language_model.model.layers.41.self_attn.q_b_proj",
338
+ "language_model.model.layers.41.self_attn.kv_a_proj_with_mqa",
339
+ "language_model.model.layers.41.self_attn.kv_b_proj",
340
+ "language_model.model.layers.41.self_attn.o_proj",
341
+ "language_model.model.layers.42.self_attn.q_a_proj",
342
+ "language_model.model.layers.42.self_attn.q_b_proj",
343
+ "language_model.model.layers.42.self_attn.kv_a_proj_with_mqa",
344
+ "language_model.model.layers.42.self_attn.kv_b_proj",
345
+ "language_model.model.layers.42.self_attn.o_proj",
346
+ "language_model.model.layers.43.self_attn.q_a_proj",
347
+ "language_model.model.layers.43.self_attn.q_b_proj",
348
+ "language_model.model.layers.43.self_attn.kv_a_proj_with_mqa",
349
+ "language_model.model.layers.43.self_attn.kv_b_proj",
350
+ "language_model.model.layers.43.self_attn.o_proj",
351
+ "language_model.model.layers.44.self_attn.q_a_proj",
352
+ "language_model.model.layers.44.self_attn.q_b_proj",
353
+ "language_model.model.layers.44.self_attn.kv_a_proj_with_mqa",
354
+ "language_model.model.layers.44.self_attn.kv_b_proj",
355
+ "language_model.model.layers.44.self_attn.o_proj",
356
+ "language_model.model.layers.45.self_attn.q_a_proj",
357
+ "language_model.model.layers.45.self_attn.q_b_proj",
358
+ "language_model.model.layers.45.self_attn.kv_a_proj_with_mqa",
359
+ "language_model.model.layers.45.self_attn.kv_b_proj",
360
+ "language_model.model.layers.45.self_attn.o_proj",
361
+ "language_model.model.layers.46.self_attn.q_a_proj",
362
+ "language_model.model.layers.46.self_attn.q_b_proj",
363
+ "language_model.model.layers.46.self_attn.kv_a_proj_with_mqa",
364
+ "language_model.model.layers.46.self_attn.kv_b_proj",
365
+ "language_model.model.layers.46.self_attn.o_proj",
366
+ "language_model.model.layers.47.self_attn.q_a_proj",
367
+ "language_model.model.layers.47.self_attn.q_b_proj",
368
+ "language_model.model.layers.47.self_attn.kv_a_proj_with_mqa",
369
+ "language_model.model.layers.47.self_attn.kv_b_proj",
370
+ "language_model.model.layers.47.self_attn.o_proj",
371
+ "language_model.model.layers.48.self_attn.q_a_proj",
372
+ "language_model.model.layers.48.self_attn.q_b_proj",
373
+ "language_model.model.layers.48.self_attn.kv_a_proj_with_mqa",
374
+ "language_model.model.layers.48.self_attn.kv_b_proj",
375
+ "language_model.model.layers.48.self_attn.o_proj",
376
+ "language_model.model.layers.49.self_attn.q_a_proj",
377
+ "language_model.model.layers.49.self_attn.q_b_proj",
378
+ "language_model.model.layers.49.self_attn.kv_a_proj_with_mqa",
379
+ "language_model.model.layers.49.self_attn.kv_b_proj",
380
+ "language_model.model.layers.49.self_attn.o_proj",
381
+ "language_model.model.layers.50.self_attn.q_a_proj",
382
+ "language_model.model.layers.50.self_attn.q_b_proj",
383
+ "language_model.model.layers.50.self_attn.kv_a_proj_with_mqa",
384
+ "language_model.model.layers.50.self_attn.kv_b_proj",
385
+ "language_model.model.layers.50.self_attn.o_proj",
386
+ "language_model.model.layers.51.self_attn.q_a_proj",
387
+ "language_model.model.layers.51.self_attn.q_b_proj",
388
+ "language_model.model.layers.51.self_attn.kv_a_proj_with_mqa",
389
+ "language_model.model.layers.51.self_attn.kv_b_proj",
390
+ "language_model.model.layers.51.self_attn.o_proj",
391
+ "language_model.model.layers.52.self_attn.q_a_proj",
392
+ "language_model.model.layers.52.self_attn.q_b_proj",
393
+ "language_model.model.layers.52.self_attn.kv_a_proj_with_mqa",
394
+ "language_model.model.layers.52.self_attn.kv_b_proj",
395
+ "language_model.model.layers.52.self_attn.o_proj",
396
+ "language_model.model.layers.53.self_attn.q_a_proj",
397
+ "language_model.model.layers.53.self_attn.q_b_proj",
398
+ "language_model.model.layers.53.self_attn.kv_a_proj_with_mqa",
399
+ "language_model.model.layers.53.self_attn.kv_b_proj",
400
+ "language_model.model.layers.53.self_attn.o_proj",
401
+ "language_model.model.layers.54.self_attn.q_a_proj",
402
+ "language_model.model.layers.54.self_attn.q_b_proj",
403
+ "language_model.model.layers.54.self_attn.kv_a_proj_with_mqa",
404
+ "language_model.model.layers.54.self_attn.kv_b_proj",
405
+ "language_model.model.layers.54.self_attn.o_proj",
406
+ "language_model.model.layers.55.self_attn.q_a_proj",
407
+ "language_model.model.layers.55.self_attn.q_b_proj",
408
+ "language_model.model.layers.55.self_attn.kv_a_proj_with_mqa",
409
+ "language_model.model.layers.55.self_attn.kv_b_proj",
410
+ "language_model.model.layers.55.self_attn.o_proj",
411
+ "language_model.model.layers.56.self_attn.q_a_proj",
412
+ "language_model.model.layers.56.self_attn.q_b_proj",
413
+ "language_model.model.layers.56.self_attn.kv_a_proj_with_mqa",
414
+ "language_model.model.layers.56.self_attn.kv_b_proj",
415
+ "language_model.model.layers.56.self_attn.o_proj",
416
+ "language_model.model.layers.57.self_attn.q_a_proj",
417
+ "language_model.model.layers.57.self_attn.q_b_proj",
418
+ "language_model.model.layers.57.self_attn.kv_a_proj_with_mqa",
419
+ "language_model.model.layers.57.self_attn.kv_b_proj",
420
+ "language_model.model.layers.57.self_attn.o_proj",
421
+ "language_model.model.layers.58.self_attn.q_a_proj",
422
+ "language_model.model.layers.58.self_attn.q_b_proj",
423
+ "language_model.model.layers.58.self_attn.kv_a_proj_with_mqa",
424
+ "language_model.model.layers.58.self_attn.kv_b_proj",
425
+ "language_model.model.layers.58.self_attn.o_proj",
426
+ "language_model.model.layers.59.self_attn.q_a_proj",
427
+ "language_model.model.layers.59.self_attn.q_b_proj",
428
+ "language_model.model.layers.59.self_attn.kv_a_proj_with_mqa",
429
+ "language_model.model.layers.59.self_attn.kv_b_proj",
430
+ "language_model.model.layers.59.self_attn.o_proj",
431
+ "language_model.model.layers.60.self_attn.q_a_proj",
432
+ "language_model.model.layers.60.self_attn.q_b_proj",
433
+ "language_model.model.layers.60.self_attn.kv_a_proj_with_mqa",
434
+ "language_model.model.layers.60.self_attn.kv_b_proj",
435
+ "language_model.model.layers.60.self_attn.o_proj",
436
+ "language_model.lm_head"
437
+ ],
438
+ "export": {
439
+ "kv_cache_group": [],
440
+ "min_kv_scale": 0.0,
441
+ "pack_method": "reorder",
442
+ "weight_format": "real_quantized",
443
+ "weight_merge_groups": null
444
+ },
445
+ "global_quant_config": {
446
+ "bias": null,
447
+ "input_tensors": [
448
+ {
449
+ "block_size": null,
450
+ "ch_axis": -1,
451
+ "dtype": "fp4",
452
+ "enable_buffer_reuse": false,
453
+ "group_size": 16,
454
+ "is_dynamic": true,
455
+ "is_scale_quant": false,
456
+ "max_input_numel": 4194304,
457
+ "mx_element_dtype": null,
458
+ "observer_cls": "PerBlockMXObserver",
459
+ "qscheme": "per_group",
460
+ "round_method": "half_even",
461
+ "scale_calculation_mode": null,
462
+ "scale_format": "float32",
463
+ "scale_type": "float32",
464
+ "symmetric": null
465
+ },
466
+ {
467
+ "block_size": null,
468
+ "ch_axis": null,
469
+ "dtype": "fp8_e4m3",
470
+ "enable_buffer_reuse": false,
471
+ "group_size": null,
472
+ "is_dynamic": false,
473
+ "is_scale_quant": true,
474
+ "max_input_numel": 4194304,
475
+ "mx_element_dtype": null,
476
+ "observer_cls": "PerTensorMinMaxObserver",
477
+ "qscheme": "per_tensor",
478
+ "round_method": "half_even",
479
+ "scale_calculation_mode": null,
480
+ "scale_format": null,
481
+ "scale_type": "float32",
482
+ "symmetric": true
483
+ }
484
+ ],
485
+ "output_tensors": null,
486
+ "target_device": null,
487
+ "weight": [
488
+ {
489
+ "block_size": null,
490
+ "ch_axis": -1,
491
+ "dtype": "fp4",
492
+ "enable_buffer_reuse": false,
493
+ "group_size": 16,
494
+ "is_dynamic": false,
495
+ "is_scale_quant": false,
496
+ "max_input_numel": 4194304,
497
+ "mx_element_dtype": null,
498
+ "observer_cls": "PerBlockMXObserver",
499
+ "qscheme": "per_group",
500
+ "round_method": "half_even",
501
+ "scale_calculation_mode": null,
502
+ "scale_format": "float32",
503
+ "scale_type": "float32",
504
+ "symmetric": null
505
+ },
506
+ {
507
+ "block_size": null,
508
+ "ch_axis": null,
509
+ "dtype": "fp8_e4m3",
510
+ "enable_buffer_reuse": false,
511
+ "group_size": null,
512
+ "is_dynamic": false,
513
+ "is_scale_quant": true,
514
+ "max_input_numel": 4194304,
515
+ "mx_element_dtype": null,
516
+ "observer_cls": "PerTensorMinMaxObserver",
517
+ "qscheme": "per_tensor",
518
+ "round_method": "half_even",
519
+ "scale_calculation_mode": null,
520
+ "scale_format": null,
521
+ "scale_type": "float32",
522
+ "symmetric": true
523
+ }
524
+ ]
525
+ },
526
+ "kv_cache_post_rope": false,
527
+ "kv_cache_quant_config": {},
528
+ "layer_quant_config": {},
529
+ "layer_type_quant_config": {},
530
+ "quant_method": "quark",
531
+ "quant_mode": "eager_mode",
532
+ "softmax_quant_spec": null,
533
+ "version": "0.12+13caa7b4951"
534
+ },
535
+ "text_config": {
536
+ "_name_or_path": "",
537
+ "add_cross_attention": false,
538
+ "architectures": [
539
+ "DeepseekV3ForCausalLM"
540
+ ],
541
+ "attention_bias": false,
542
+ "attention_dropout": 0.0,
543
+ "auto_map": {
544
+ "AutoConfig": "configuration_deepseek.DeepseekV3Config",
545
+ "AutoModel": "modeling_deepseek.DeepseekV3Model",
546
+ "AutoModelForCausalLM": "modeling_deepseek.DeepseekV3ForCausalLM"
547
+ },
548
+ "aux_loss_alpha": 0.001,
549
+ "bad_words_ids": null,
550
+ "begin_suppress_tokens": null,
551
+ "bos_token_id": 163584,
552
+ "chunk_size_feed_forward": 0,
553
+ "cross_attention_hidden_size": null,
554
+ "decoder_start_token_id": null,
555
+ "diversity_penalty": 0.0,
556
+ "do_sample": false,
557
+ "dtype": "bfloat16",
558
+ "early_stopping": false,
559
+ "encoder_no_repeat_ngram_size": 0,
560
+ "eos_token_id": 163586,
561
+ "ep_size": 1,
562
+ "exponential_decay_length_penalty": null,
563
+ "finetuning_task": null,
564
+ "first_k_dense_replace": 1,
565
+ "forced_bos_token_id": null,
566
+ "forced_eos_token_id": null,
567
+ "hidden_act": "silu",
568
+ "hidden_size": 7168,
569
+ "id2label": {
570
+ "0": "LABEL_0",
571
+ "1": "LABEL_1"
572
+ },
573
+ "initializer_range": 0.02,
574
+ "intermediate_size": 18432,
575
+ "is_decoder": false,
576
+ "is_encoder_decoder": false,
577
+ "kv_lora_rank": 512,
578
+ "label2id": {
579
+ "LABEL_0": 0,
580
+ "LABEL_1": 1
581
+ },
582
+ "length_penalty": 1.0,
583
+ "max_length": 20,
584
+ "max_position_embeddings": 262144,
585
+ "min_length": 0,
586
+ "model_type": "deepseek_v3",
587
+ "moe_intermediate_size": 2048,
588
+ "moe_layer_freq": 1,
589
+ "n_group": 1,
590
+ "n_routed_experts": 384,
591
+ "n_shared_experts": 1,
592
+ "no_repeat_ngram_size": 0,
593
+ "norm_topk_prob": true,
594
+ "num_attention_heads": 64,
595
+ "num_beam_groups": 1,
596
+ "num_beams": 1,
597
+ "num_experts_per_tok": 8,
598
+ "num_hidden_layers": 61,
599
+ "num_key_value_heads": 64,
600
+ "num_nextn_predict_layers": 0,
601
+ "num_return_sequences": 1,
602
+ "output_attentions": false,
603
+ "output_hidden_states": false,
604
+ "output_scores": false,
605
+ "pad_token_id": 163839,
606
+ "prefix": null,
607
+ "pretraining_tp": 1,
608
+ "problem_type": null,
609
+ "pruned_heads": {},
610
+ "q_lora_rank": 1536,
611
+ "qk_nope_head_dim": 128,
612
+ "qk_rope_head_dim": 64,
613
+ "quantization_config": {
614
+ "config_groups": {
615
+ "group_0": {
616
+ "input_activations": null,
617
+ "output_activations": null,
618
+ "targets": [
619
+ "Linear"
620
+ ],
621
+ "weights": {
622
+ "actorder": null,
623
+ "block_structure": null,
624
+ "dynamic": false,
625
+ "group_size": 32,
626
+ "num_bits": 4,
627
+ "observer": "minmax",
628
+ "observer_kwargs": {},
629
+ "strategy": "group",
630
+ "symmetric": true,
631
+ "type": "int"
632
+ }
633
+ }
634
+ },
635
+ "format": "pack-quantized",
636
+ "ignore": [
637
+ "re:.*self_attn.*",
638
+ "re:.*shared_experts.*",
639
+ "re:.*mlp\\.(gate|up|gate_up|down)_proj.*",
640
+ "re:.*lm_head.*",
641
+ "re:vision_tower.*",
642
+ "re:mm_projector.*"
643
+ ],
644
+ "kv_cache_scheme": null,
645
+ "quant_method": "compressed-tensors",
646
+ "quantization_status": "compressed"
647
+ },
648
+ "remove_invalid_values": false,
649
+ "repetition_penalty": 1.0,
650
+ "return_dict": true,
651
+ "return_dict_in_generate": false,
652
+ "rms_norm_eps": 1e-05,
653
+ "rope_scaling": {
654
+ "beta_fast": 32.0,
655
+ "beta_slow": 1.0,
656
+ "factor": 64.0,
657
+ "mscale": 1.0,
658
+ "mscale_all_dim": 1.0,
659
+ "original_max_position_embeddings": 4096,
660
+ "type": "yarn"
661
+ },
662
+ "rope_theta": 50000.0,
663
+ "routed_scaling_factor": 2.827,
664
+ "scoring_func": "sigmoid",
665
+ "sep_token_id": null,
666
+ "seq_aux": true,
667
+ "suppress_tokens": null,
668
+ "task_specific_params": null,
669
+ "temperature": 1.0,
670
+ "tf_legacy_loss": false,
671
+ "tie_encoder_decoder": false,
672
+ "tie_word_embeddings": false,
673
+ "tokenizer_class": null,
674
+ "top_k": 50,
675
+ "top_p": 1.0,
676
+ "topk_group": 1,
677
+ "topk_method": "noaux_tc",
678
+ "torchscript": false,
679
+ "typical_p": 1.0,
680
+ "use_bfloat16": false,
681
+ "use_cache": true,
682
+ "v_head_dim": 128,
683
+ "vocab_size": 163840
684
+ },
685
+ "tie_word_embeddings": false,
686
+ "transformers_version": "4.57.6",
687
+ "use_unified_vision_chunk": true,
688
+ "video_placeholder": "<|kimi_k25_video_placeholder|>",
689
+ "vision_config": {
690
+ "init_pos_emb_height": 64,
691
+ "init_pos_emb_time": 4,
692
+ "init_pos_emb_width": 64,
693
+ "merge_kernel_size": [
694
+ 2,
695
+ 2
696
+ ],
697
+ "merge_type": "sd2_tpool",
698
+ "mm_hidden_size": 1152,
699
+ "mm_projector_type": "patchmerger",
700
+ "model_type": "",
701
+ "patch_size": 14,
702
+ "pos_emb_type": "divided_fixed",
703
+ "projector_hidden_act": "gelu",
704
+ "projector_ln_eps": 1e-05,
705
+ "text_hidden_size": 7168,
706
+ "video_attn_type": "spatial_temporal",
707
+ "vt_hidden_size": 1152,
708
+ "vt_intermediate_size": 4304,
709
+ "vt_num_attention_heads": 16,
710
+ "vt_num_hidden_layers": 27
711
+ }
712
+ }
configuration_deepseek.py ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copy from https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/main/configuration_deepseek.py
2
+
3
+ from transformers.configuration_utils import PretrainedConfig
4
+ from transformers.utils import logging
5
+
6
+ logger = logging.get_logger(__name__)
7
+
8
+ DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
9
+
10
+
11
+ class DeepseekV3Config(PretrainedConfig):
12
+ r"""
13
+ This is the configuration class to store the configuration of a [`DeepseekV3Model`]. It is used to instantiate an DeepSeek
14
+ model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
15
+ defaults will yield a similar configuration to that of the DeepSeek-V3.
16
+
17
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
18
+ documentation from [`PretrainedConfig`] for more information.
19
+
20
+
21
+ Args:
22
+ vocab_size (`int`, *optional*, defaults to 129280):
23
+ Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
24
+ `inputs_ids` passed when calling [`DeepseekV3Model`]
25
+ hidden_size (`int`, *optional*, defaults to 4096):
26
+ Dimension of the hidden representations.
27
+ intermediate_size (`int`, *optional*, defaults to 11008):
28
+ Dimension of the MLP representations.
29
+ moe_intermediate_size (`int`, *optional*, defaults to 1407):
30
+ Dimension of the MoE representations.
31
+ num_hidden_layers (`int`, *optional*, defaults to 32):
32
+ Number of hidden layers in the Transformer decoder.
33
+ num_nextn_predict_layers (`int`, *optional*, defaults to 1):
34
+ Number of nextn predict layers in the DeepSeekV3 Model.
35
+ num_attention_heads (`int`, *optional*, defaults to 32):
36
+ Number of attention heads for each attention layer in the Transformer decoder.
37
+ n_shared_experts (`int`, *optional*, defaults to None):
38
+ Number of shared experts, None means dense model.
39
+ n_routed_experts (`int`, *optional*, defaults to None):
40
+ Number of routed experts, None means dense model.
41
+ routed_scaling_factor (`float`, *optional*, defaults to 1.0):
42
+ Scaling factor or routed experts.
43
+ topk_method (`str`, *optional*, defaults to `gready`):
44
+ Topk method used in routed gate.
45
+ n_group (`int`, *optional*, defaults to None):
46
+ Number of groups for routed experts.
47
+ topk_group (`int`, *optional*, defaults to None):
48
+ Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
49
+ num_experts_per_tok (`int`, *optional*, defaults to None):
50
+ Number of selected experts, None means dense model.
51
+ moe_layer_freq (`int`, *optional*, defaults to 1):
52
+ The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
53
+ first_k_dense_replace (`int`, *optional*, defaults to 0):
54
+ Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
55
+ \--k dense layers--/
56
+ norm_topk_prob (`bool`, *optional*, defaults to False):
57
+ Whether to normalize the weights of the routed experts.
58
+ scoring_func (`str`, *optional*, defaults to 'softmax'):
59
+ Method of computing expert weights.
60
+ aux_loss_alpha (`float`, *optional*, defaults to 0.001):
61
+ Auxiliary loss weight coefficient.
62
+ seq_aux = (`bool`, *optional*, defaults to True):
63
+ Whether to compute the auxiliary loss for each individual sample.
64
+ num_key_value_heads (`int`, *optional*):
65
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
66
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
67
+ `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
68
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
69
+ by meanpooling all the original heads within that group. For more details checkout [this
70
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
71
+ `num_attention_heads`.
72
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
73
+ The non-linear activation function (function or string) in the decoder.
74
+ max_position_embeddings (`int`, *optional*, defaults to 2048):
75
+ The maximum sequence length that this model might ever be used with.
76
+ initializer_range (`float`, *optional*, defaults to 0.02):
77
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
78
+ rms_norm_eps (`float`, *optional*, defaults to 1e-06):
79
+ The epsilon used by the rms normalization layers.
80
+ use_cache (`bool`, *optional*, defaults to `True`):
81
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
82
+ relevant if `config.is_decoder=True`.
83
+ pad_token_id (`int`, *optional*):
84
+ Padding token id.
85
+ bos_token_id (`int`, *optional*, defaults to 1):
86
+ Beginning of stream token id.
87
+ eos_token_id (`int`, *optional*, defaults to 2):
88
+ End of stream token id.
89
+ pretraining_tp (`int`, *optional*, defaults to 1):
90
+ Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
91
+ document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
92
+ necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
93
+ issue](https://github.com/pytorch/pytorch/issues/76232).
94
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
95
+ Whether to tie weight embeddings
96
+ rope_theta (`float`, *optional*, defaults to 10000.0):
97
+ The base period of the RoPE embeddings.
98
+ rope_scaling (`Dict`, *optional*):
99
+ Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
100
+ strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
101
+ `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
102
+ `max_position_embeddings` to the expected new maximum.
103
+ attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
104
+ Whether to use a bias in the query, key, value and output projection layers during self-attention.
105
+ attention_dropout (`float`, *optional*, defaults to 0.0):
106
+ The dropout ratio for the attention probabilities.
107
+
108
+ ```python
109
+ >>> from transformers import DeepseekV3Model, DeepseekV3Config
110
+
111
+ >>> # Initializing a Deepseek-V3 style configuration
112
+ >>> configuration = DeepseekV3Config()
113
+
114
+ >>> # Accessing the model configuration
115
+ >>> configuration = model.config
116
+ ```"""
117
+
118
+ model_type = "deepseek_v3"
119
+ keys_to_ignore_at_inference = ["past_key_values"]
120
+
121
+ def __init__(
122
+ self,
123
+ vocab_size=129280,
124
+ hidden_size=7168,
125
+ intermediate_size=18432,
126
+ moe_intermediate_size=2048,
127
+ num_hidden_layers=61,
128
+ num_nextn_predict_layers=1,
129
+ num_attention_heads=128,
130
+ num_key_value_heads=128,
131
+ n_shared_experts=1,
132
+ n_routed_experts=256,
133
+ ep_size=1,
134
+ routed_scaling_factor=2.5,
135
+ kv_lora_rank=512,
136
+ q_lora_rank=1536,
137
+ qk_rope_head_dim=64,
138
+ v_head_dim=128,
139
+ qk_nope_head_dim=128,
140
+ topk_method='noaux_tc',
141
+ n_group=8,
142
+ topk_group=4,
143
+ num_experts_per_tok=8,
144
+ moe_layer_freq=1,
145
+ first_k_dense_replace=3,
146
+ norm_topk_prob=True,
147
+ scoring_func='sigmoid',
148
+ aux_loss_alpha=0.001,
149
+ seq_aux=True,
150
+ hidden_act="silu",
151
+ max_position_embeddings=4096,
152
+ initializer_range=0.02,
153
+ rms_norm_eps=1e-6,
154
+ use_cache=True,
155
+ pad_token_id=None,
156
+ bos_token_id=0,
157
+ eos_token_id=1,
158
+ pretraining_tp=1,
159
+ tie_word_embeddings=False,
160
+ rope_theta=10000.0,
161
+ rope_scaling=None,
162
+ attention_bias=False,
163
+ attention_dropout=0.0,
164
+ **kwargs,
165
+ ):
166
+ self.vocab_size = vocab_size
167
+ self.max_position_embeddings = max_position_embeddings
168
+ self.hidden_size = hidden_size
169
+ self.intermediate_size = intermediate_size
170
+ self.moe_intermediate_size = moe_intermediate_size
171
+ self.num_hidden_layers = num_hidden_layers
172
+ self.num_nextn_predict_layers = num_nextn_predict_layers
173
+ self.num_attention_heads = num_attention_heads
174
+ self.n_shared_experts = n_shared_experts
175
+ self.n_routed_experts = n_routed_experts
176
+ self.ep_size = ep_size
177
+ self.routed_scaling_factor = routed_scaling_factor
178
+ self.kv_lora_rank = kv_lora_rank
179
+ self.q_lora_rank = q_lora_rank
180
+ self.qk_rope_head_dim = qk_rope_head_dim
181
+ self.v_head_dim = v_head_dim
182
+ self.qk_nope_head_dim = qk_nope_head_dim
183
+ self.topk_method = topk_method
184
+ self.n_group = n_group
185
+ self.topk_group = topk_group
186
+ self.num_experts_per_tok = num_experts_per_tok
187
+ self.moe_layer_freq = moe_layer_freq
188
+ self.first_k_dense_replace = first_k_dense_replace
189
+ self.norm_topk_prob = norm_topk_prob
190
+ self.scoring_func = scoring_func
191
+ self.aux_loss_alpha = aux_loss_alpha
192
+ self.seq_aux = seq_aux
193
+ # for backward compatibility
194
+ if num_key_value_heads is None:
195
+ num_key_value_heads = num_attention_heads
196
+
197
+ self.num_key_value_heads = num_key_value_heads
198
+ self.hidden_act = hidden_act
199
+ self.initializer_range = initializer_range
200
+ self.rms_norm_eps = rms_norm_eps
201
+ self.pretraining_tp = pretraining_tp
202
+ self.use_cache = use_cache
203
+ self.rope_theta = rope_theta
204
+ self.rope_scaling = rope_scaling
205
+ self.attention_bias = attention_bias
206
+ self.attention_dropout = attention_dropout
207
+
208
+ super().__init__(
209
+ pad_token_id=pad_token_id,
210
+ bos_token_id=bos_token_id,
211
+ eos_token_id=eos_token_id,
212
+ tie_word_embeddings=tie_word_embeddings,
213
+ **kwargs,
214
+ )
configuration_kimi_k25.py ADDED
@@ -0,0 +1,123 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from transformers.configuration_utils import PretrainedConfig
2
+
3
+ try:
4
+ from configuration_deepseek import DeepseekV3Config
5
+ except ImportError:
6
+ from .configuration_deepseek import DeepseekV3Config
7
+
8
+
9
+ class KimiK25VisionConfig(PretrainedConfig):
10
+
11
+ def __init__(
12
+ self,
13
+ patch_size: int = 14,
14
+ init_pos_emb_height: int = 64,
15
+ init_pos_emb_width: int = 64,
16
+ init_pos_emb_time: int = 4,
17
+ pos_emb_type: str = 'divided_fixed',
18
+ vt_num_attention_heads: int = 16,
19
+ vt_num_hidden_layers: int = 27,
20
+ vt_hidden_size: int = 1152,
21
+ vt_intermediate_size: int = 4304,
22
+ merge_kernel_size: tuple = (2, 2),
23
+ video_attn_type: str = 'spatial_temporal',
24
+ merge_type: str = 'sd2_tpool',
25
+ _attn_implementation: str = 'flash_attention_2',
26
+ # MM Projector parameters
27
+ mm_projector_type: str = 'patchmerger',
28
+ mm_hidden_size: int | None = None,
29
+ projector_hidden_act: str = "gelu",
30
+ projector_ln_eps: float = 1e-5,
31
+ # Other parameters
32
+ ignore_index: int = -100,
33
+ media_placeholder_token_id: int = 163605,
34
+ pad_token_id: int = 0,
35
+ use_unified_vision_chunk: bool = True,
36
+ video_placeholder="<|kimi_k25_video_placeholder|>",
37
+ text_hidden_size=7168,
38
+ **vision_config_kwargs):
39
+
40
+ self.patch_size = patch_size
41
+ self.init_pos_emb_height = init_pos_emb_height
42
+ self.init_pos_emb_width = init_pos_emb_width
43
+ self.init_pos_emb_time = init_pos_emb_time
44
+ self.pos_emb_type = pos_emb_type
45
+ self.vt_num_attention_heads = vt_num_attention_heads
46
+ self.vt_num_hidden_layers = vt_num_hidden_layers
47
+ self.vt_hidden_size = vt_hidden_size
48
+ self.vt_intermediate_size = vt_intermediate_size
49
+ self.merge_kernel_size = merge_kernel_size
50
+ self.video_attn_type = video_attn_type
51
+ self.merge_type = merge_type
52
+ self._attn_implementation = _attn_implementation
53
+
54
+ # MM Projector config
55
+ self.mm_projector_type = mm_projector_type
56
+ self.mm_hidden_size = mm_hidden_size if mm_hidden_size is not None else vt_hidden_size
57
+ self.projector_hidden_act = projector_hidden_act
58
+ self.projector_ln_eps = projector_ln_eps
59
+ self.text_hidden_size = text_hidden_size
60
+
61
+
62
+ class KimiK25Config(PretrainedConfig):
63
+ """Kimi-K2.5 model configuration.
64
+
65
+ Args:
66
+ text_config (dict | DeepseekV3Config): Configuration for the text model.
67
+
68
+ Vision Tower Parameters (from MoonViT3dConfig):
69
+ patch_size (int): Patch size for vision tower.
70
+ init_pos_emb_height (int): Initial position embedding height.
71
+ init_pos_emb_width (int): Initial position embedding width.
72
+ init_pos_emb_time (int): Initial position embedding time dimension.
73
+ pos_emb_type (str): Type of position embedding.
74
+ vt_num_attention_heads (int): Number of attention heads in vision tower.
75
+ vt_num_hidden_layers (int): Number of hidden layers in vision tower.
76
+ vt_hidden_size (int): Hidden size of vision tower.
77
+ vt_intermediate_size (int): Intermediate size in vision tower FFN.
78
+ merge_kernel_size (tuple): Kernel size for patch merging.
79
+ video_attn_type (str): Type of video attention.
80
+ merge_type (str): Type of merge operation.
81
+ _attn_implementation (str): Attention implementation type.
82
+
83
+ MM Projector Parameters (from MultiModalProjectorConfig):
84
+ mm_projector_type (str): Type of multimodal projector.
85
+ mm_hidden_size (int): Hidden size from vision tower (should match vt_hidden_size).
86
+ projector_hidden_act (str): Activation function for projector.
87
+ projector_ln_eps (float): Layer norm epsilon for projector.
88
+
89
+ Other Parameters:
90
+ ignore_index (int): The ignore index for the loss function.
91
+ media_placeholder_token_id (int): The token ID to use for media placeholders.
92
+ pad_token_id (int): The token ID to use for padding.
93
+ """
94
+
95
+ model_type = "kimi_k25"
96
+
97
+ def __init__(
98
+ self,
99
+ text_config: dict | DeepseekV3Config = None,
100
+ vision_config: dict | KimiK25VisionConfig = None,
101
+ # Other parameters
102
+ ignore_index: int = -100,
103
+ media_placeholder_token_id: int = 163605,
104
+ pad_token_id: int = 0,
105
+ use_unified_vision_chunk: bool = True,
106
+ video_placeholder="<|kimi_k25_video_placeholder|>",
107
+ **kwargs,
108
+ ):
109
+ if isinstance(text_config, dict):
110
+ text_config = DeepseekV3Config(**text_config)
111
+ if isinstance(vision_config, dict):
112
+ vision_config = KimiK25VisionConfig(**vision_config)
113
+ self.text_config = text_config
114
+ self.vision_config = vision_config
115
+ # Other config
116
+ self.ignore_index = ignore_index
117
+ self.media_placeholder_token_id = media_placeholder_token_id
118
+ self.use_unified_vision_chunk = use_unified_vision_chunk
119
+ self.video_placeholder = video_placeholder
120
+ if getattr(self.text_config, "quantization_config", None) is not None:
121
+ self.quantization_config = self.text_config.quantization_config
122
+
123
+ super().__init__(pad_token_id=pad_token_id, **kwargs)
generation_config.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 163584,
4
+ "do_sample": true,
5
+ "eos_token_id": 163586,
6
+ "pad_token_id": 163839,
7
+ "transformers_version": "4.57.6"
8
+ }
kimi_k25_vision_processing.py ADDED
@@ -0,0 +1,251 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Image processor class for Kimi-K2.5.
2
+ """
3
+
4
+ import json
5
+ from typing import Any, Dict, Optional, Union
6
+
7
+ import numpy as np
8
+ import torch
9
+ from PIL import Image
10
+ from transformers.image_processing_utils import (BaseImageProcessor,
11
+ BatchFeature)
12
+ from transformers.utils import TensorType
13
+
14
+ from .media_utils import (MediaInput, VideoChunkInput, _to_tensor,
15
+ ensure_media_type, get_video_meta, image_to_np,
16
+ navit_patchify, navit_resize_image,
17
+ navit_resize_video, normalize,
18
+ real_sample_fps_and_max_num_frames, timestamp_as_str)
19
+
20
+ try:
21
+ from mecord import VideoReader
22
+ except ImportError:
23
+ VideoReader = None
24
+
25
+
26
+ def resampling(video_bytes: bytes,
27
+ sample_indices: list[int],
28
+ key_indices=None,
29
+ frame_time_info=None,
30
+ num_threads=4) -> str:
31
+ video = VideoReader(video_bytes,
32
+ num_threads=num_threads,
33
+ frame_time_info=frame_time_info,
34
+ key_indices=key_indices)
35
+ # extract target frames
36
+ frames = video[sample_indices]
37
+ frames = [Image.fromarray(frame) for frame in frames]
38
+ return frames
39
+
40
+
41
+ class KimiK25VisionProcessor(BaseImageProcessor):
42
+ model_type = "kimi_k25"
43
+
44
+ def __init__(
45
+ self,
46
+ media_proc_cfg: dict,
47
+ **kwargs,
48
+ ):
49
+ super().__init__(**kwargs)
50
+ self.media_proc_cfg = media_proc_cfg
51
+ self.num_frames_per_chunk = media_proc_cfg[
52
+ 'temporal_merge_kernel_size']
53
+
54
+ def media_tokens_calculator(self, media: MediaInput):
55
+ media = ensure_media_type(media)
56
+ ret = self.get_resize_config(media)
57
+ return ret['num_tokens']
58
+
59
+ @classmethod
60
+ def make_chunk_prompt(cls, timestamp_text: str) -> str:
61
+ return f"{timestamp_text}<|media_begin|>video<|media_content|><|media_pad|><|media_end|>"
62
+
63
+ def split_video_chunks(self,
64
+ video_url: str | bytes) -> list[list[Image.Image]]:
65
+ # video_url should be base64 str or bytes
66
+ video_spec = get_video_meta(video_url)
67
+ sample_fps = min(self.media_proc_cfg['sample_fps'], video_spec.fps)
68
+ sampled_nframes = max(
69
+ round(video_spec.num_frames * sample_fps / video_spec.fps), 1)
70
+ frame_inds = np.linspace(0, video_spec.num_frames - 1,
71
+ sampled_nframes).round().astype(int)
72
+ frame_inds = frame_inds.tolist()
73
+ sampled_frame_ids = []
74
+ temporal_merge_kernel_size = self.media_proc_cfg[
75
+ "temporal_merge_kernel_size"]
76
+ num_chunks = 0
77
+ chunk_timestamp = []
78
+ for i in range(0, len(frame_inds), temporal_merge_kernel_size):
79
+ sampled_frame_ids.extend(frame_inds[i:i +
80
+ temporal_merge_kernel_size])
81
+ start_time = frame_inds[i] / float(video_spec.fps)
82
+ timestamp_text = timestamp_as_str(
83
+ start_time, self.media_proc_cfg["timestamp_mode"])
84
+ chunk_timestamp.append(timestamp_text)
85
+ num_chunks += 1
86
+
87
+ sampled_frames = resampling(video_url, sampled_frame_ids)
88
+ chunks = []
89
+ for chunk_id in range(num_chunks):
90
+ chunk = sampled_frames[chunk_id *
91
+ temporal_merge_kernel_size:(chunk_id + 1) *
92
+ temporal_merge_kernel_size]
93
+ chunks.append(
94
+ VideoChunkInput(type="video_chunk",
95
+ video_chunk=chunk,
96
+ prompt=self.make_chunk_prompt(
97
+ chunk_timestamp[chunk_id])))
98
+ return chunks
99
+
100
+ def get_resize_config(self, media_input: MediaInput) -> dict:
101
+ if media_input['type'] == 'image':
102
+ w, h = media_input['image'].size
103
+ ret = navit_resize_image(
104
+ w, h, self.media_proc_cfg['patch_size'],
105
+ self.media_proc_cfg['merge_kernel_size'],
106
+ self.media_proc_cfg['in_patch_limit'],
107
+ self.media_proc_cfg['patch_limit_on_one_side'],
108
+ self.media_proc_cfg['fixed_output_tokens'])
109
+ return ret
110
+ elif media_input['type'] == 'video_chunk':
111
+ frame = media_input['video_chunk'][0]
112
+ width, height = frame.size
113
+ num_frames = len(media_input["video_chunk"])
114
+ fps = 1.0
115
+
116
+ sample_fps, max_num_frames_each_video = real_sample_fps_and_max_num_frames(
117
+ media_input["type"],
118
+ self.media_proc_cfg['sample_fps'],
119
+ self.media_proc_cfg['max_num_frames_each_video'],
120
+ )
121
+
122
+ in_patch_limit_each_frame = self.media_proc_cfg[
123
+ 'in_patch_limit_each_frame']
124
+ if in_patch_limit_each_frame is None:
125
+ in_patch_limit_each_frame = self.media_proc_cfg[
126
+ 'in_patch_limit']
127
+
128
+ ret = navit_resize_video(
129
+ width,
130
+ height,
131
+ num_frames,
132
+ fps,
133
+ sample_fps,
134
+ self.media_proc_cfg['patch_size'],
135
+ self.media_proc_cfg['merge_kernel_size'],
136
+ in_patch_limit_each_frame,
137
+ self.media_proc_cfg['patch_limit_on_one_side'],
138
+ self.media_proc_cfg['in_patch_limit_video'],
139
+ max_num_frames_each_video,
140
+ self.media_proc_cfg['fixed_output_tokens'],
141
+ )
142
+ return ret
143
+ else:
144
+ raise ValueError("Unsupported type: {}".format(
145
+ media_input['type']))
146
+
147
+ def resize_image(self, image: Image.Image, new_width: int, new_height: int,
148
+ pad_width: int, pad_height: int) -> np.ndarray:
149
+ image_np = image_to_np(image, (new_width, new_height), "resize")
150
+ image_np = np.pad(
151
+ image_np,
152
+ ((0, pad_height), (0, pad_width), (0, 0)),
153
+ mode="constant",
154
+ constant_values=0,
155
+ )
156
+ return image_np
157
+
158
+ def preprocess(
159
+ self,
160
+ medias: list[MediaInput],
161
+ return_tensors: Optional[Union[str, TensorType]] = None,
162
+ ) -> BatchFeature:
163
+ """
164
+ Preprocess a atom vision input (images/video_chunk) into model-ready tensors.
165
+
166
+ Args:
167
+ medias: List of MediaInput.
168
+ return_tensors: Desired output format ('pt', 'np', 'tf', or None).
169
+
170
+ Returns:
171
+ BatchFeature containing 'pixel_values' and 'grid_thws' tensors.
172
+ """
173
+ if not isinstance(medias, list):
174
+ medias = [medias]
175
+ if medias:
176
+ pixel_values = []
177
+ for item in medias:
178
+ item = ensure_media_type(item)
179
+ resize_config = self.get_resize_config(item)
180
+ new_width, new_height, pad_width, pad_height = resize_config[
181
+ 'new_width'], resize_config['new_height'], resize_config[
182
+ 'pad_width'], resize_config['pad_height']
183
+ if item['type'] == 'image':
184
+ image = item['image']
185
+ image_np = self.resize_image(image, new_width, new_height,
186
+ pad_width, pad_height)
187
+ pixel_values.append(np.expand_dims(image_np, axis=0))
188
+ elif item['type'] == 'video_chunk':
189
+ pixels = []
190
+ for frame in item['video_chunk']:
191
+ frame_np = self.resize_image(frame, new_width,
192
+ new_height, pad_width,
193
+ pad_height)
194
+ pixels.append(frame_np)
195
+ pixel_values.append(np.stack(pixels, axis=0))
196
+ else:
197
+ raise ValueError("Unsupported type: {}".format(
198
+ item['type']))
199
+ normalized_pixel_values = []
200
+ image_std_inv = 1.0 / np.array(self.media_proc_cfg['image_std'])
201
+ image_mean = np.array(self.media_proc_cfg['image_mean'])
202
+ for pixels in pixel_values:
203
+ pixels = normalize(pixels, image_mean, image_std_inv)
204
+ pixels_and_thw = navit_patchify(
205
+ pixels,
206
+ self.media_proc_cfg['patch_size'],
207
+ )
208
+ normalized_pixel_values.append(pixels_and_thw)
209
+
210
+ pixel_values = torch.cat([
211
+ _to_tensor(pixel_value['pixel_values'])
212
+ for pixel_value in normalized_pixel_values
213
+ ])
214
+ grid_thws = torch.cat([
215
+ _to_tensor(pixel_value['grid_thw'],
216
+ dtype=torch.int64).unsqueeze(0)
217
+ for pixel_value in normalized_pixel_values
218
+ ])
219
+
220
+ data = {
221
+ 'pixel_values': pixel_values,
222
+ 'grid_thws': grid_thws,
223
+ }
224
+
225
+ else:
226
+ data = {}
227
+
228
+ return BatchFeature(data=data, tensor_type=return_tensors)
229
+
230
+ def __repr__(self):
231
+ return f"KimiK25VisionProcessor(media_proc_cfg={self.media_proc_cfg})"
232
+
233
+ def to_dict(self) -> Dict[str, Any]:
234
+ output = super().to_dict()
235
+ output["media_proc_cfg"] = self.media_proc_cfg
236
+ if "media_processor" in output:
237
+ del output["media_processor"]
238
+ return output
239
+
240
+ @classmethod
241
+ def from_dict(cls, config_dict: Dict[str, Any], **kwargs):
242
+ config = config_dict.copy()
243
+ media_proc_cfg = config.pop("media_proc_cfg", {})
244
+ return cls(media_proc_cfg=media_proc_cfg, **config, **kwargs)
245
+
246
+ def to_json_string(self):
247
+ dictionary = self.to_dict()
248
+ for key, value in dictionary.items():
249
+ if hasattr(value, 'tolist'):
250
+ dictionary[key] = value.tolist()
251
+ return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"
media_utils.py ADDED
@@ -0,0 +1,368 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import base64
2
+ import io
3
+ import math
4
+ import os
5
+ from datetime import datetime, timezone
6
+ from typing import List, Literal, Optional, TypedDict
7
+
8
+ import numpy as np
9
+ from PIL import Image
10
+ from pydantic import BaseModel, Field
11
+
12
+ try:
13
+ from mecord import VideoReader
14
+ except ImportError:
15
+ VideoReader = None
16
+
17
+
18
+ class VideoSpec(BaseModel):
19
+ media_type: str = Literal['video']
20
+ height: int = Field(..., gt=0, description="video frame height")
21
+ width: int = Field(..., gt=0, description="video frame width")
22
+ num_frames: int = Field(..., gt=0, description="num frames")
23
+ fps: float = Field(..., gt=0, description="average fps")
24
+
25
+ # optional, help to accelerate video reading
26
+ key_indices: list[int] = Field(None, description="key indices")
27
+ frame_time_info: dict = Field(None, description="frame time info")
28
+
29
+
30
+ class ImageInput(TypedDict):
31
+ type: Literal['image']
32
+ image: Image.Image
33
+
34
+
35
+ class VideoChunkInput(TypedDict):
36
+ type: Literal['video_chunk']
37
+ video_chunk: List[Image.Image]
38
+ prompt: Optional[str] = None
39
+
40
+
41
+ MediaInput = ImageInput | VideoChunkInput
42
+
43
+
44
+ def get_video_meta(video_src: bytes | str | os.PathLike,
45
+ accurate: bool = True) -> dict:
46
+ """Get the dimensions of a video."""
47
+ if isinstance(video_src, os.PathLike):
48
+ video_src = str(video_src)
49
+ # if b64 string, decode to bytes
50
+ if isinstance(video_src,
51
+ str) and video_src.startswith('data:video/mp4;base64,'):
52
+ video_src = base64.b64decode(video_src.split(',')[1])
53
+ video = VideoReader(video_src, auto_init=accurate, num_threads=1)
54
+ assert video.num_frames > 0, "Invalid video format."
55
+ assert video.original_width > 0 and video.original_height > 0, (
56
+ "Invalid video format.")
57
+ assert video.avg_fps > 0, "Invalid video format."
58
+ return VideoSpec(media_type='video',
59
+ height=video.original_height,
60
+ width=video.original_width,
61
+ num_frames=video.num_frames,
62
+ fps=video.avg_fps,
63
+ key_indices=video.key_indices,
64
+ frame_time_info=video.frame_time_info)
65
+
66
+
67
+ def timestamp_as_str(timestamp: float,
68
+ timestamp_mode: str = "hh:mm:ss.fff") -> str:
69
+ """Convert a timestamp to a string in the format of HH:MM:SS.mmm."""
70
+ if timestamp_mode == "hh:mm:ss.fff":
71
+ return (datetime.fromtimestamp(timestamp,
72
+ tz=timezone.utc).strftime("%H:%M:%S") +
73
+ f".{int((timestamp % 1) * 1000):03d}")
74
+ elif timestamp_mode == "mm:ss.fff":
75
+ return (datetime.fromtimestamp(timestamp,
76
+ tz=timezone.utc).strftime("%M:%S") +
77
+ f".{int((timestamp % 1) * 1000):03d}")
78
+ elif timestamp_mode == "mm:ss":
79
+ return datetime.fromtimestamp(timestamp,
80
+ tz=timezone.utc).strftime("%M:%S")
81
+ else:
82
+ raise ValueError(f"Invalid timestamp mode: {timestamp_mode}")
83
+
84
+
85
+ def navit_resize_image(
86
+ width: int,
87
+ height: int,
88
+ patch_size: int,
89
+ merge_kernel_size: int,
90
+ in_patch_limit: int,
91
+ patch_limit_on_one_side: int,
92
+ fixed_output_tokens: int | None,
93
+ ):
94
+ # Apply the patch limits.
95
+ s1 = math.sqrt(
96
+ in_patch_limit /
97
+ (max(1.0, width // patch_size) * max(1.0, height // patch_size)))
98
+ s2 = patch_limit_on_one_side * patch_size / width
99
+ s3 = patch_limit_on_one_side * patch_size / height
100
+ scale = min(1.0, s1, s2, s3)
101
+ new_w, new_h = max(1, int(width * scale)), max(1, int(height * scale))
102
+ new_w = min(new_w, patch_limit_on_one_side * patch_size)
103
+ new_h = min(new_h, patch_limit_on_one_side * patch_size)
104
+
105
+ # Calculate the padding to make the height and width divisible by the merge kernel size and patch size.
106
+ factor = merge_kernel_size * patch_size
107
+
108
+ pad_height = (factor - new_h % factor) % factor
109
+ pad_width = (factor - new_w % factor) % factor
110
+
111
+ if fixed_output_tokens is not None:
112
+ num_tokens = fixed_output_tokens
113
+ else:
114
+ # Calculate new dimensions after padding and patching
115
+ token_height = (new_h + pad_height) // factor
116
+ token_width = (new_w + pad_width) // factor
117
+
118
+ assert token_height * merge_kernel_size <= patch_limit_on_one_side, (
119
+ f"token_height {token_height} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
120
+ )
121
+ assert token_width * merge_kernel_size <= patch_limit_on_one_side, (
122
+ f"token_width {token_width} * merge_kernel_size {merge_kernel_size} > patch_limit_on_one_side {patch_limit_on_one_side}"
123
+ )
124
+
125
+ num_tokens = token_height * token_width
126
+ return {
127
+ "num_tokens": num_tokens,
128
+ "new_width": new_w,
129
+ "new_height": new_h,
130
+ "pad_width": pad_width,
131
+ "pad_height": pad_height,
132
+ "sampled_nframes": 1,
133
+ }
134
+
135
+
136
+ def navit_resize_video(
137
+ width: int,
138
+ height: int,
139
+ nframes: int,
140
+ avg_fps: float,
141
+ sample_fps: float,
142
+ patch_size: int,
143
+ merge_kernel_size: int,
144
+ in_patch_limit_each_frame: int,
145
+ patch_limit_on_one_side: int,
146
+ in_patch_limit_total: int | None,
147
+ max_num_frames_each_video: int | None,
148
+ fixed_output_tokens_each_frame: int | None,
149
+ ):
150
+ sample_fps = min(sample_fps, avg_fps)
151
+ # Calculate the number of frames to sample based on target FPS
152
+ sampled_nframes = max(round(nframes * sample_fps / avg_fps), 1)
153
+ if max_num_frames_each_video is not None:
154
+ sampled_nframes = min(sampled_nframes, max_num_frames_each_video)
155
+
156
+ if in_patch_limit_total is not None:
157
+ in_patch_limit_each_frame = min(
158
+ round(in_patch_limit_total / sampled_nframes),
159
+ in_patch_limit_each_frame)
160
+
161
+ ret = navit_resize_image(
162
+ width,
163
+ height,
164
+ patch_size,
165
+ merge_kernel_size,
166
+ in_patch_limit_each_frame,
167
+ patch_limit_on_one_side,
168
+ fixed_output_tokens_each_frame,
169
+ )
170
+ ret["sampled_nframes"] = sampled_nframes
171
+ return ret
172
+
173
+
174
+ def real_sample_fps_and_max_num_frames(
175
+ type_name: Literal["video", "video_chunk"],
176
+ sample_fps: float,
177
+ max_num_frames_each_video: int | None,
178
+ ) -> tuple[int, int | None]:
179
+ if type_name == "video":
180
+ return sample_fps, max_num_frames_each_video
181
+ elif type_name == "video_chunk":
182
+ max_num_frames_each_video = None
183
+ sample_fps = math.inf
184
+ return sample_fps, max_num_frames_each_video
185
+ else:
186
+ return math.inf, None
187
+
188
+
189
+ def _to_pil(data: str | bytes):
190
+ if isinstance(data, Image.Image):
191
+
192
+ return data.convert("RGB")
193
+ elif isinstance(data, str):
194
+ if data.startswith("data:"):
195
+ raw_base64 = data.split(",")[1]
196
+ return Image.open(io.BytesIO(
197
+ base64.b64decode(raw_base64))).convert("RGB")
198
+ else:
199
+ return Image.open(data).convert("RGB")
200
+ elif isinstance(data, bytes):
201
+ return Image.open(io.BytesIO(data)).convert("RGB")
202
+ else:
203
+ raise ValueError(f"Unsupported data type: {type(data)}")
204
+
205
+
206
+ def ensure_media_type(media: MediaInput) -> MediaInput:
207
+ if media['type'] == 'image':
208
+ media['image'] = _to_pil(media['image'])
209
+ return media
210
+ elif media['type'] == 'video_chunk':
211
+ media['video_chunk'] = [
212
+ _to_pil(frame) for frame in media['video_chunk']
213
+ ]
214
+ return media
215
+ else:
216
+ raise ValueError(f"Unsupported media type: {media['type']}")
217
+
218
+
219
+ def image_to_np(
220
+ image: Image.Image,
221
+ resize_to: tuple[int, int] | None = None,
222
+ mode: str = "resize",
223
+ raise_error_for_ill_resize: bool = True,
224
+ ) -> np.ndarray:
225
+ """Convert an image to a numpy array.
226
+
227
+ Args:
228
+ content: The image to convert.
229
+ resize_to: The size to resize the image to.
230
+ mode: The mode to resize the image to.
231
+ raise_error_for_ill_resize: Whether to raise an error for ill-sized resize.
232
+
233
+ Returns:
234
+ A numpy array.
235
+ """
236
+ assert isinstance(image, Image.Image), "image must be a PIL Image"
237
+ if resize_to is not None:
238
+ if mode == "resize":
239
+ image = image.resize(resize_to, resample=Image.Resampling.BICUBIC)
240
+
241
+ elif mode == "rescale_and_pad_to_center":
242
+ scale = min(resize_to[0] / image.width,
243
+ resize_to[1] / image.height, 1.0)
244
+ new_width = round(image.width * scale)
245
+ new_height = round(image.height * scale)
246
+ if new_width == 0 or new_height == 0:
247
+ if raise_error_for_ill_resize:
248
+ raise ValueError(
249
+ f"Invalid resize to: {resize_to}, from image size: {image.size}"
250
+ )
251
+ else:
252
+ return np.zeros((resize_to[1], resize_to[0], 3),
253
+ dtype=np.uint8)
254
+
255
+ image = image.resize((new_width, new_height),
256
+ resample=Image.Resampling.BICUBIC)
257
+ padding_left = (resize_to[0] - new_width) // 2
258
+ padding_right = resize_to[0] - new_width - padding_left
259
+ padding_top = (resize_to[1] - new_height) // 2
260
+ padding_bottom = resize_to[1] - new_height - padding_top
261
+ image = np.asarray(image)
262
+ image = np.pad(
263
+ image,
264
+ ((padding_top, padding_bottom), (padding_left, padding_right),
265
+ (0, 0)),
266
+ mode="constant",
267
+ constant_values=0,
268
+ )
269
+ assert image.shape == (resize_to[1], resize_to[0], 3)
270
+
271
+ elif mode == "rescale_and_pad_to_rightbottom":
272
+ scale = min(resize_to[0] / image.width,
273
+ resize_to[1] / image.height, 1.0)
274
+ new_width = round(image.width * scale)
275
+ new_height = round(image.height * scale)
276
+ if new_width == 0 or new_height == 0:
277
+ if raise_error_for_ill_resize:
278
+ raise ValueError(
279
+ f"Invalid resize to: {resize_to}, from image size: {image.size}"
280
+ )
281
+ else:
282
+ return np.zeros((resize_to[1], resize_to[0], 3),
283
+ dtype=np.uint8)
284
+
285
+ image = image.resize((new_width, new_height),
286
+ resample=Image.Resampling.BICUBIC)
287
+ padding_right = resize_to[0] - new_width
288
+ padding_bottom = resize_to[1] - new_height
289
+ image = np.asarray(image)
290
+ image = np.pad(
291
+ image,
292
+ ((0, padding_bottom), (0, padding_right), (0, 0)),
293
+ mode="constant",
294
+ constant_values=0,
295
+ )
296
+ assert image.shape == (resize_to[1], resize_to[0], 3)
297
+
298
+ else:
299
+ raise ValueError(f"Invalid mode: {mode}")
300
+
301
+ if isinstance(image, Image.Image):
302
+ return np.asarray(image)
303
+ else:
304
+ return image
305
+
306
+
307
+ def navit_patchify(pixel_values: np.ndarray,
308
+ patch_size: int) -> dict[str, np.ndarray]:
309
+ """Reshape the pixel values to a navit shape.
310
+
311
+ Args:
312
+ pixel_values: np.ndarray, shape (t, h, w, c)
313
+ patch_size: int
314
+
315
+ Returns:
316
+ dict[str, np.ndarray]
317
+ - patches: np.ndarray, shape (t * h//patch_size * w//patch_size, c, patch_size, patch_size)
318
+ - grid_thw: np.ndarray, (t, h//patch_size, w//patch_size)
319
+ """
320
+ T, H, W, C = pixel_values.shape
321
+ assert C == 3, "pixel_values must have 3 channels"
322
+
323
+ patches = pixel_values.reshape(T, H // patch_size, patch_size,
324
+ W // patch_size, patch_size, C)
325
+ # (T, H//patch_size, W//patch_size, C, patch_size, patch_size)
326
+ patches = patches.transpose(0, 1, 3, 5, 2, 4)
327
+ patches = patches.reshape(-1, C, patch_size, patch_size)
328
+ grid_thw = np.array([T, H // patch_size, W // patch_size])
329
+ return {"pixel_values": patches, "grid_thw": grid_thw}
330
+
331
+
332
+ def normalize(x: np.ndarray,
333
+ mean,
334
+ std_inv,
335
+ pixels_dtype: np.dtype = np.float32) -> np.ndarray:
336
+ """Normalize the image.
337
+
338
+ Args:
339
+ x: The image to normalize. The shape is (..., 3). The dtype is uint8. The range is [0, 255].
340
+ mean: The mean of the image.
341
+ std_inv: The inverse of the std of the image.
342
+ pixels_dtype: The dtype of the image.
343
+ Returns:
344
+ The normalized image. The shape is (..., 3). The dtype is determined by the pixels_dtype.
345
+ """
346
+ x = (x / 255.0).astype(pixels_dtype)
347
+ x -= mean
348
+ x *= std_inv
349
+ return x
350
+
351
+
352
+ def _to_tensor(data, **kwargs):
353
+ import torch
354
+
355
+ if isinstance(data, np.ndarray):
356
+ return torch.from_numpy(data).to(**kwargs)
357
+ elif isinstance(data, torch.Tensor):
358
+ return data.to(**kwargs)
359
+ elif isinstance(data, list):
360
+ return [_to_tensor(item, **kwargs) for item in data]
361
+ elif isinstance(data, tuple):
362
+ return tuple(_to_tensor(item, **kwargs) for item in data)
363
+ elif isinstance(data, dict):
364
+ return {k: _to_tensor(v, **kwargs) for k, v in data.items()}
365
+ elif data is None:
366
+ return None
367
+ else:
368
+ raise ValueError(f"Unsupported data type: {type(data)}")
model-00001-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:77a75493828a842fcd9c14b1fbe0e2f811e8bedd13ab6110b87b15b7e59cca36
3
+ size 4999509288
model-00002-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba77afe8bdea9b3b4c6cfff792ce062c6c433d9d56164c0a1738da5b581bffa7
3
+ size 4997054960
model-00003-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:44700ce97e57a4be9cde2fe465e618c49277bb0c96443e78c444c1521da881a0
3
+ size 4997469344
model-00004-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bacbe527c826512b97dee20294c2b332c538df2f14bfb343648694dc5b585892
3
+ size 4996137176
model-00005-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d391829b01425f828d218854a79740782b84fd2b5ebefecfa780a55005666de
3
+ size 4997469224
model-00006-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed0570bd83a3546fc5a76616c042c58223ec41774f74ab6a8a09c18198847f4e
3
+ size 4996137296
model-00007-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f9fa7a02dfdd0ecc7cf82d5c75c9b5bec02b61a83a565a8e47ae9f44b8e4fe10
3
+ size 4997469104
model-00008-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5ccb14300375810dcbe5c2dffdf99103f954a37bcfbce62efb8110b4f08d3325
3
+ size 4996137416
model-00009-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:906fe773fbeca2831b2210de9b9bef64f59de4e938be94af1d15cdb7cb3f1da1
3
+ size 4997468984
model-00010-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5ec3a0d946552f9576b6979dd9d4146fa5fa730a41f3dbfcd70b77f55f4dd90d
3
+ size 4996137536
model-00011-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1992bafdf6fa282a4b6c2643f46c06d005a9a3dcefb53044aff41095be753488
3
+ size 4997468864
model-00012-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b35f6e0e7e32567c6bbd39a4e5ce3ad598650bea51b7288828ae2e20c653a98
3
+ size 4996137656
model-00013-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:325cfd11c11ea2f7ff145981693c525f491f47b727e7eb529350d5503a1a23f4
3
+ size 4997468784
model-00014-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a96bbe18bf39e357b22d81b6178e621f80fc045128d425c3066b362db84c2d53
3
+ size 4996137736
model-00015-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ed0dae6587962c48288f5c30c0a6c09524c2da1e2a71c37fd95a0b12fc48b8b2
3
+ size 4997468784
model-00016-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:75aa9590824da3a20f3a7f69c7b10f4fb1e67266f7a65a98c8bc0edf84a77226
3
+ size 4996137736
model-00017-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c2f7f156ffa4602d249c38d9bd26545e19d6a68eb75c9f255aa21d91a29a42a8
3
+ size 4997468784
model-00018-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0999734434ecc135c5ba476b49c69589b6705ae6ed2f80279f8eb1b425df48ee
3
+ size 4996137736
model-00019-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:91fa659e9e8ad089870406fb5494dc104678d1e3b2ebafd961e24597ceb7e388
3
+ size 4997470416
model-00020-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:666b21f70ba1db493d5ffd06af5432a060426ffd274981334901e7ad32928182
3
+ size 4996140152
model-00021-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5a8f84c86eb471563f3d3d51341fe57ce7cfc1b3f6c3f63d96c358f0d8793812
3
+ size 4997471120
model-00022-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8beda06991a3170cf23ff47ad753048cb232b69fc3cf856a79ad856d8848f3c1
3
+ size 4996140152
model-00023-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb398fe67cda3af7318075b083985a2c205685ae2a793788a667bf5ac38dee3a
3
+ size 4997471120
model-00024-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c9f11714ce369eb45147601dd18e3cfca3628fd6dd37a463c8aa97c16262dc21
3
+ size 4996140152
model-00025-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bfe35fac0631da59bf4f957002daf26d3ef80db935e296944e450c91c37054f5
3
+ size 4997471120
model-00026-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:791f434c80418f3c5a2da183dfec447f2197abd35f95697b00122b0fec7b287b
3
+ size 4996140152
model-00027-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a19997a36fd7bce89b24d10c45fc65e08ea6a9bd88d77024c8161ffbfdb83717
3
+ size 4997471120
model-00028-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4f80a3fe8b608db353be55085c1fd407173b409884c0b4c25dea196f5c6d0691
3
+ size 4996140152
model-00029-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1ea061972fd9b8f89154b3c3c945843e66ba18fd04514f5442dbc6abf8385c29
3
+ size 4997471120
model-00030-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d2fb46c1ae8e6773ad141cc99fbf1fef9dfc0ae76c57948e1b2be5968569de04
3
+ size 4993417088
model-00031-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f8d9b4e433c8eed1ab2d8474dd52c6b3bc54ec4d3c9ec6f626aaa32a0144efc2
3
+ size 5000194200
model-00032-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ffc91671dd32146bfb2fe0a7f97c662ef5342cf8c88af2e796841fb0274cf4e7
3
+ size 4997472400
model-00033-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:87a0fdf8e1404a70502fed5ce9323f5d28bcadf9291c9edc68d15303d4b7b906
3
+ size 4996138872
model-00034-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3c295bed3493264aa43ed7d3f9c6a71d03cc3bb9a9a6aaef29c4cc9bb820c2d8
3
+ size 4997472184
model-00035-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b899fffb7cb106dcb75c666348cdbac292f2c1e3a88bfd039e8392a914d43e90
3
+ size 4996139088
model-00036-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fac4e2d921754fc66ad87dd329483291de4cca79e0cd0d17d5fd6302a6561b25
3
+ size 4997472064
model-00037-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:49bb8b65c1f032256c5e50dd52be56dee2c28655e77e99eb4c01de0231489217
3
+ size 4996139208
model-00038-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6dd7ad07083bbcef3a8064f7605c3b1eb995739835351792beb298d4b4414323
3
+ size 4997471944
model-00039-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52f16d0bae9c9388fd4f2092cba1aa9482981f4a36e6a55ef71958ba3545bcf2
3
+ size 4996139328
model-00040-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:180a6248a182e9d48ff2abd39b8bbd57c53b4a3dfddc3f177af5ddb5b2a0c2c1
3
+ size 4997471824
model-00041-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d6dea05436fd657b83c05c57971db86cbab60f6798e64f3b6ab8cab4e13dc73c
3
+ size 4996139448
model-00042-of-00119.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d2ed16abb8835b8b21ccccacf5b2b7707ab5e80b57f2cb0dc0a5a4dad58157bd
3
+ size 4997471704