# tested on transformers==4.57.1
import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiny-random/kimi-linear"
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="cuda",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)

messages = [
    {"role": "system", "content": "You are a helpful assistant provided by Moonshot-AI."},
    {"role": "user", "content": "Is 123 a prime?"}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    tokenize=True,
).to(model.device)
print(input_ids)
generated_ids = model.generate(inputs=input_ids, max_new_tokens=500)
response = tokenizer.batch_decode(generated_ids)[0]
print(response)

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-Linear-48B-A3B-Instruct"
save_folder = "/tmp/tiny-random/kimi-linear"

Path(save_folder).mkdir(parents=True, exist_ok=True)
tokenizer = AutoTokenizer.from_pretrained(
    source_model_id, trust_remote_code=True)
tokenizer.save_pretrained(save_folder)
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({
    "head_dim": 32,
    "hidden_size": 8,
    "intermediate_size": 32,
    "linear_attn_config": {
        "full_attn_layers": [4],
        "head_dim": 32,
        "kda_layers": [1, 2, 3],
        "num_heads": 8,
        "short_conv_kernel_size": 4,
    },
    "num_attention_heads": 8,
    "num_key_value_heads": 8,
    "moe_intermediate_size": 32,
    "num_hidden_layers": 5,
})
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()
n_parms = 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, 0.1)
        print(name, p.shape, (p.numel() / n_parms * 100), '%')
model.save_pretrained(save_folder)

with open(f"{save_folder}/config.json", "r", encoding='utf-8') as f:
    config_json = json.load(f)
    config_json['auto_map'] = {k: f'{source_model_id}--' + 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()

Printing the model:

text

KimiLinearForCausalLM(
  (model): KimiLinearModel(
    (embed_tokens): Embedding(163840, 8, padding_idx=163839)
    (layers): ModuleList(
      (0): KimiDecoderLayer(
        (self_attn): KimiDeltaAttention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (k_proj): Linear(in_features=8, out_features=256, bias=False)
          (v_proj): Linear(in_features=8, out_features=256, bias=False)
          (q_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (k_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (v_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (f_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (f_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (b_proj): Linear(in_features=8, out_features=8, bias=False)
          (g_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (g_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (o_norm): FusedRMSNormGated(32, eps=1e-05, activation=sigmoid)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (mlp): KimiMLP(
          (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): KimiRMSNorm()
        (post_attention_layernorm): KimiRMSNorm()
      )
      (1-2): 2 x KimiDecoderLayer(
        (self_attn): KimiDeltaAttention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (k_proj): Linear(in_features=8, out_features=256, bias=False)
          (v_proj): Linear(in_features=8, out_features=256, bias=False)
          (q_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (k_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (v_conv1d): ShortConvolution(256, 256, kernel_size=(4,), stride=(1,), padding=(3,), groups=256, bias=False, activation=silu, backend=triton)
          (f_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (f_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (b_proj): Linear(in_features=8, out_features=8, bias=False)
          (g_a_proj): Linear(in_features=8, out_features=32, bias=False)
          (g_b_proj): Linear(in_features=32, out_features=256, bias=False)
          (o_norm): FusedRMSNormGated(32, eps=1e-05, activation=sigmoid)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
        )
        (block_sparse_moe): KimiSparseMoeBlock(
          (experts): ModuleList(
            (0-255): 256 x KimiBlockSparseMLP(
              (w1): Linear(in_features=8, out_features=32, bias=False)
              (w2): Linear(in_features=32, out_features=8, bias=False)
              (w3): Linear(in_features=8, out_features=32, bias=False)
              (act_fn): SiLUActivation()
            )
          )
          (gate): KimiMoEGate()
          (shared_experts): KimiMLP(
            (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): KimiRMSNorm()
        (post_attention_layernorm): KimiRMSNorm()
      )
      (3-4): 2 x KimiDecoderLayer(
        (self_attn): KimiMLAAttention(
          (q_proj): Linear(in_features=8, out_features=1536, bias=False)
          (kv_a_proj_with_mqa): Linear(in_features=8, out_features=576, bias=False)
          (kv_a_layernorm): KimiRMSNorm()
          (kv_b_proj): Linear(in_features=512, out_features=2048, bias=False)
          (o_proj): Linear(in_features=1024, out_features=8, bias=False)
        )
        (block_sparse_moe): KimiSparseMoeBlock(
          (experts): ModuleList(
            (0-255): 256 x KimiBlockSparseMLP(
              (w1): Linear(in_features=8, out_features=32, bias=False)
              (w2): Linear(in_features=32, out_features=8, bias=False)
              (w3): Linear(in_features=8, out_features=32, bias=False)
              (act_fn): SiLUActivation()
            )
          )
          (gate): KimiMoEGate()
          (shared_experts): KimiMLP(
            (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): KimiRMSNorm()
        (post_attention_layernorm): KimiRMSNorm()
      )
    )
    (norm): KimiRMSNorm()
  )
  (lm_head): Linear(in_features=8, out_features=163840, bias=False)
)