tiny-random/kimi-k2.6

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from moonshotai/Kimi-K2.6.

File path	Size
model.safetensors	9.0MB

Example usage:

• vLLM

vllm serve tiny-random/kimi-k2.6 --trust-remote-code

• Transformers

import base64
import requests
import torch
from transformers import AutoModel, AutoProcessor

model_id = "tiny-random/kimi-k2.6"
image_url = "https://avatars.githubusercontent.com/u/0"
image_base64 = base64.b64encode(requests.get(image_url).content).decode()
messages = [
    {
        'role': 'user',
        'content': [
            {'type': 'text', 'text': 'Describe this image in detail.'},
            {
                'type': 'image',
                'image_url': f'data:image/png;base64,{image_base64}',
            },
        ],
    }
]
processor = AutoProcessor.from_pretrained(
    model_id,
    trust_remote_code=True,
)
model = AutoModel.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="cuda" if torch.cuda.is_available() else "cpu",
    trust_remote_code=True,
).eval()

# Text generation is not compatible with the latest version of transformers (v5.5)
# so we only show a dummy model forward step here
inputs = processor(
    messages=messages,
    tokenize=False,
    return_tensors="pt"
).to(model.device)
inputs.input_ids[0, -1] = model.config.media_placeholder_token_id
print(inputs.keys())
result = model(**inputs)
print(result)

Codes to create this repo:

Click to expand

import json
from pathlib import Path

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

source_model_id = "moonshotai/Kimi-K2.6"
save_folder = "/tmp/tiny-random/kimi-k26"

Path(save_folder).mkdir(parents=True, exist_ok=True)

suffixes = ['.json', '.py', '.model', '.jinja']
for f in list_repo_files(source_model_id, repo_type="model"):
    if any(f.endswith(suffix) for suffix in suffixes) and not f.endswith('.index.json'):
        hf_hub_download(
            repo_id=source_model_id,
            filename=f,
            repo_type="model",
            local_dir=save_folder
        )

def replace_file(filepath, old_string, new_string):
    with open(filepath, 'r', encoding='utf-8') as f:
        code = f.read()
    code = code.replace(old_string, new_string)
    with open(filepath, 'w', encoding='utf-8') as f:
        f.write(code)

replace_file(f'{save_folder}/configuration_kimi_k25.py',
             "from configuration_deepseek import DeepseekV3Config",
             "from transformers import DeepseekV3Config")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "from .modeling_deepseek import DeepseekV3ForCausalLM",
             "from transformers import DeepseekV3ForCausalLM")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "use_deterministic_attn=self.use_deterministic_attn",
             "")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "def tie_weights(self):",
             "def tie_weights(self, *args, **kwargs):")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "_supports_flash_attn_2 = True",
             "_supports_flash_attn_2 = True\n    _supports_flash_attn = True")
with open(f'{save_folder}/config.json') as f:
    config_json = json.load(f)

config_json['text_config'].update({
    'first_k_dense_replace': 1,
    'num_hidden_layers': 2,
    'hidden_size': 8,
    'intermediate_size': 32,
    'moe_intermediate_size': 32,
    # 'n_routed_experts': 32,
    # 'n_shared_experts': 1,
    'num_attention_heads': 4,
    # 'num_experts_per_tok': 8,
    'num_key_value_heads': 4,
    'q_lora_rank': 32,
    # 'qk_nope_head_dim': 64,
    # 'qk_rope_head_dim': 192,
    # 'v_head_dim': 64,
    'tie_word_embeddings': False,
})
del config_json['text_config']['quantization_config']
config_json['vision_config'].update({
    'mm_hidden_size': 64,
    'text_hidden_size': 8,
    'vt_hidden_size': 64,
    'vt_intermediate_size': 128,
    'vt_num_attention_heads': 2,
    'vt_num_hidden_layers': 2,
})
config_json['vision_config']['_attn_implementation'] = 'eager'
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 = AutoModel.from_config(config, trust_remote_code=True, attn_implementation='eager')
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()
num_params = sum(p.numel() for p in model.parameters())
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.2)
        print(name, p.shape, p.dtype, p.device, f'{p.numel() / num_params * 100: .2f}%')
model.save_pretrained(save_folder)
replace_file(f'{save_folder}/configuration_kimi_k25.py',
             "from configuration_deepseek import DeepseekV3Config",
             "from transformers import DeepseekV3Config")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "from .modeling_deepseek import DeepseekV3ForCausalLM",
             "from transformers import DeepseekV3ForCausalLM")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "use_deterministic_attn=self.use_deterministic_attn",
             "")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "def tie_weights(self):",
             "def tie_weights(self, *args, **kwargs):")
replace_file(f'{save_folder}/modeling_kimi_k25.py',
             "_supports_flash_attn_2 = True",
             "_supports_flash_attn_2 = True\n    _supports_flash_attn = True")

Printing the model:

Click to expand

KimiK25ForConditionalGeneration(
  (vision_tower): MoonViT3dPretrainedModel(
    (patch_embed): MoonVision3dPatchEmbed(
      (proj): Conv2d(3, 64, kernel_size=(14, 14), stride=(14, 14))
      (pos_emb): Learnable2DInterpPosEmbDivided_fixed()
    )
    (encoder): MoonViT3dEncoder(
      (rope_2d): Rope2DPosEmbRepeated(dim=32, max_height=512, max_width=512, theta_base=10000)
      (blocks): ModuleList(
        (0-1): 2 x MoonViTEncoderLayer(
          (norm0): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          (norm1): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
          (mlp): MLP2(
            (fc0): Linear(in_features=64, out_features=128, bias=True)
            (fc1): Linear(in_features=128, out_features=64, bias=True)
            (activation): GELUTanh()
          )
          (wqkv): Linear(in_features=64, out_features=192, bias=True)
          (wo): Linear(in_features=64, out_features=64, bias=True)
        )
      )
      (final_layernorm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
    )
  )
  (mm_projector): PatchMergerMLP(
    (pre_norm): LayerNorm((64,), eps=1e-05, elementwise_affine=True)
    (proj): Sequential(
      (0): Linear(in_features=256, out_features=256, bias=True)
      (1): GELU(approximate='none')
      (2): Linear(in_features=256, out_features=8, bias=True)
    )
  )
  (language_model): DeepseekV3ForCausalLM(
    (model): DeepseekV3Model(
      (embed_tokens): Embedding(163840, 8, padding_idx=163839)
      (layers): ModuleList(
        (0): DeepseekV3DecoderLayer(
          (self_attn): DeepseekV3Attention(
            (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
            (q_a_layernorm): DeepseekV3RMSNorm((32,), eps=1e-06)
            (q_b_proj): Linear(in_features=32, out_features=768, bias=False)
            (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
            (kv_a_layernorm): DeepseekV3RMSNorm((512,), eps=1e-06)
            (kv_b_proj): Linear(in_features=512, out_features=1024, bias=False)
            (o_proj): Linear(in_features=512, out_features=8, bias=False)
          )
          (mlp): DeepseekV3MLP(
            (gate_proj): Linear(in_features=8, out_features=32, bias=False)
            (up_proj): Linear(in_features=8, out_features=32, bias=False)
            (down_proj): Linear(in_features=32, out_features=8, bias=False)
            (act_fn): SiLUActivation()
          )
          (input_layernorm): DeepseekV3RMSNorm((8,), eps=1e-05)
          (post_attention_layernorm): DeepseekV3RMSNorm((8,), eps=1e-05)
        )
        (1): DeepseekV3DecoderLayer(
          (self_attn): DeepseekV3Attention(
            (q_a_proj): Linear(in_features=8, out_features=32, bias=False)
            (q_a_layernorm): DeepseekV3RMSNorm((32,), eps=1e-06)
            (q_b_proj): Linear(in_features=32, out_features=768, bias=False)
            (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
            (kv_a_layernorm): DeepseekV3RMSNorm((512,), eps=1e-06)
            (kv_b_proj): Linear(in_features=512, out_features=1024, bias=False)
            (o_proj): Linear(in_features=512, out_features=8, bias=False)
          )
          (mlp): DeepseekV3MoE(
            (experts): DeepseekV3NaiveMoe(
              (act_fn): SiLUActivation()
            )
            (gate): DeepseekV3TopkRouter()
            (shared_experts): DeepseekV3MLP(
              (gate_proj): Linear(in_features=8, out_features=32, bias=False)
              (up_proj): Linear(in_features=8, out_features=32, bias=False)
              (down_proj): Linear(in_features=32, out_features=8, bias=False)
              (act_fn): SiLUActivation()
            )
          )
          (input_layernorm): DeepseekV3RMSNorm((8,), eps=1e-05)
          (post_attention_layernorm): DeepseekV3RMSNorm((8,), eps=1e-05)
        )
      )
      (norm): DeepseekV3RMSNorm((8,), eps=1e-05)
      (rotary_emb): DeepseekV3RotaryEmbedding()
    )
    (lm_head): Linear(in_features=8, out_features=163840, bias=False)
  )
)

Test environment:

• torch: 2.11.0
• transformers: 5.5.0