File size: 7,680 Bytes
93c8028
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1bec50b
93c8028
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1bec50b
93c8028
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
---
library_name: transformers
pipeline_tag: text-generation
inference: true
widget:
  - text: Hello!
    example_title: Hello world
    group: Python
base_model:
- moonshotai/Kimi-K2-Instruct
---

This tiny model is for debugging. It is randomly initialized with the config adapted from [moonshotai/Kimi-K2-Instruct](https://huggingface.co/moonshotai/Kimi-K2-Instruct).

### Example usage:

- vLLM

```bash
vllm serve tiny-random/kimi-k2 --trust-remote-code
```

- Transformers

```python
import torch
import transformers

model_id = "tiny-random/kimi-k2"
pipe = transformers.pipelines.pipeline(
    'text-generation',
    model=model_id,
    trust_remote_code=True,
    device_map='cuda',
    torch_dtype=torch.bfloat16,
)
messages = [
    {"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
    {"role": "user", "content": [{"type": "text", "text": "Please give a brief self-introduction."}]},
]
print(pipe(messages, max_new_tokens=128, do_sample=True, temperature=0.7, top_p=0.95))
```

### Codes to create this repo:

```python
import json
from pathlib import Path

import accelerate
import torch
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
    AutoConfig,
    AutoModelForCausalLM,
    AutoProcessor,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "moonshotai/Kimi-K2-Instruct"
save_folder = "/tmp/tiny-random/kimi-k2"

Path(save_folder).mkdir(parents=True, exist_ok=True)
with open(hf_hub_download(source_model_id, filename='tokenizer_config.json', repo_type='model'), 'r', encoding='utf-8') as f:
    tokenizer_config_json = json.load(f)
tokenizer_config_json['auto_map']['AutoTokenizer'][0] = f'{source_model_id}--' + \
    tokenizer_config_json["auto_map"]["AutoTokenizer"][0]
with open(f"{save_folder}/tokenizer_config.json", "w", encoding='utf-8') as f:
    json.dump(tokenizer_config_json, f, indent=2)
hf_hub_download(source_model_id, filename='tiktoken.model', repo_type='model',
                local_dir=save_folder, local_dir_use_symlinks=True, cache_dir='/tmp/')

with open(hf_hub_download(source_model_id, filename='config.json', repo_type='model'), 'r', encoding='utf-8') as f:
    config_json = json.load(f)
for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'
config_json.update({
    'first_k_dense_replace': 1,
    'num_hidden_layers': 2,
    'hidden_size': 32,
    'intermediate_size': 64,
    'kv_lora_rank': 384,
    'moe_intermediate_size': 64,
    'n_routed_experts': 32,
    'n_shared_experts': 1,
    'num_attention_heads': 1,
    'num_experts_per_tok': 8,
    'num_key_value_heads': 1,
    'q_lora_rank': 32,
    'qk_nope_head_dim': 64,
    'qk_rope_head_dim': 192,  # vllm mla kernel supports 576 only, FA supports head dim <= 256
    'v_head_dim': 64,
    'tie_word_embeddings': False,
})
config_json['rope_scaling']['rope_type'] = 'yarn'
del config_json['quantization_config']
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)

config = AutoConfig.from_pretrained(
    save_folder,
    trust_remote_code=True,
)
print(config)
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
torch.set_default_dtype(torch.float32)
if file_exists(filename="generation_config.json", repo_id=source_model_id, repo_type='model'):
    model.generation_config = GenerationConfig.from_pretrained(
        source_model_id, trust_remote_code=True,
    )
set_seed(42)
model = model.cpu()  # cpu is more stable for random initialization across machines
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape)
model.save_pretrained(save_folder)
# print(model)
with open(f"{save_folder}/config.json", "r", encoding='utf-8') as f:
    config_json = json.load(f)
    config_json['auto_map'] = {k: v.split('--')[-1] for k, v in config_json['auto_map'].items()}
with open(f"{save_folder}/config.json", "w", encoding='utf-8') as f:
    json.dump(config_json, f, indent=2)
# for python_file in Path(save_folder).glob('*.py'):
#     python_file.unlink()
with open(f'{save_folder}/modeling_deepseek.py', 'r', encoding='utf-8') as f:
    codes = f.read()
codes = codes.replace(
    "past_length = past_key_values.seen_tokens",
    "past_length = past_key_values.seen_tokens if hasattr(past_key_values, 'seen_tokens') else past_key_values.get_seq_length() # fix cache api deprecation"
)
codes = codes.replace(
    "max_cache_length = past_key_values.get_max_length()",
    "max_cache_length = past_key_values.get_max_length() if hasattr(past_key_values, 'get_max_length') else past_key_values.get_max_cache_shape() # fix cache api deprecation"
)
with open(f'{save_folder}/modeling_deepseek.py', 'w', encoding='utf-8') as f:
    f.write(codes)
```

### Printing the model:

```text
DeepseekV3ForCausalLM(
  (model): DeepseekV3Model(
    (embed_tokens): Embedding(163840, 32)
    (layers): ModuleList(
      (0): DeepseekV3DecoderLayer(
        (self_attn): DeepseekV3Attention(
          (q_a_proj): Linear(in_features=32, out_features=32, bias=False)
          (q_a_layernorm): DeepseekV3RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=32, out_features=576, bias=False)
          (kv_a_layernorm): DeepseekV3RMSNorm()
          (kv_b_proj): Linear(in_features=384, out_features=128, bias=False)
          (o_proj): Linear(in_features=64, out_features=32, bias=False)
          (rotary_emb): DeepseekV3YarnRotaryEmbedding()
        )
        (mlp): DeepseekV3MLP(
          (gate_proj): Linear(in_features=32, out_features=64, bias=False)
          (up_proj): Linear(in_features=32, out_features=64, bias=False)
          (down_proj): Linear(in_features=64, out_features=32, bias=False)
          (act_fn): SiLU()
        )
        (input_layernorm): DeepseekV3RMSNorm()
        (post_attention_layernorm): DeepseekV3RMSNorm()
      )
      (1): DeepseekV3DecoderLayer(
        (self_attn): DeepseekV3Attention(
          (q_a_proj): Linear(in_features=32, out_features=32, bias=False)
          (q_a_layernorm): DeepseekV3RMSNorm()
          (q_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=32, out_features=576, bias=False)
          (kv_a_layernorm): DeepseekV3RMSNorm()
          (kv_b_proj): Linear(in_features=384, out_features=128, bias=False)
          (o_proj): Linear(in_features=64, out_features=32, bias=False)
          (rotary_emb): DeepseekV3YarnRotaryEmbedding()
        )
        (mlp): DeepseekV3MoE(
          (experts): ModuleList(
            (0-31): 32 x DeepseekV3MLP(
              (gate_proj): Linear(in_features=32, out_features=64, bias=False)
              (up_proj): Linear(in_features=32, out_features=64, bias=False)
              (down_proj): Linear(in_features=64, out_features=32, bias=False)
              (act_fn): SiLU()
            )
          )
          (gate): MoEGate()
          (shared_experts): DeepseekV3MLP(
            (gate_proj): Linear(in_features=32, out_features=64, bias=False)
            (up_proj): Linear(in_features=32, out_features=64, bias=False)
            (down_proj): Linear(in_features=64, out_features=32, bias=False)
            (act_fn): SiLU()
          )
        )
        (input_layernorm): DeepseekV3RMSNorm()
        (post_attention_layernorm): DeepseekV3RMSNorm()
      )
    )
    (norm): DeepseekV3RMSNorm()
  )
  (lm_head): Linear(in_features=32, out_features=163840, bias=False)
)
```