README.md

---
library_name: transformers
base_model:
- MiniMaxAI/MiniMax-M2
---

This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from [MiniMaxAI/MiniMax-M2](https://huggingface.co/MiniMaxAI/MiniMax-M2).

### Example usage:

- vLLM

```bash
vllm serve yujiepan/minimax-m2-tiny-random --trust-remote-code
```

- Transformers

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

model_id = "yujiepan/minimax-m2-tiny-random"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
)
pipe = pipeline('text-generation', model=model,
                tokenizer=tokenizer, trust_remote_code=True)
print(pipe('Write an article about Artificial Intelligence.'))
```

### 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,
    AutoTokenizer,
    GenerationConfig,
    set_seed,
)

source_model_id = "MiniMaxAI/MiniMax-M2"
save_folder = "/tmp/yujiepan/minimax-m2-tiny-random"

processor = AutoTokenizer.from_pretrained(source_model_id)
processor.save_pretrained(save_folder)

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)

config_json["attn_type_list"] = [1, 1]
for k, v in config_json['auto_map'].items():
    config_json['auto_map'][k] = f'{source_model_id}--{v}'
config_json['head_dim'] = 32
config_json['hidden_size'] = 8
config_json['intermediate_size'] = 64
config_json['num_attention_heads'] = 8
config_json['num_key_value_heads'] = 4
config_json['num_hidden_layers'] = 2
config_json['mlp_intermediate_size'] = 64
config_json['num_local_experts'] = 32
config_json['rotary_dim'] = 16
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)
automap = config_json['auto_map']
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
torch.set_default_dtype(torch.float32)
# according to source model, gat is in FP32
for i in range(config.num_hidden_layers):
    model.model.layers[i].block_sparse_moe.gate.float()
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()
with torch.no_grad():
    for name, p in sorted(model.named_parameters()):
        torch.nn.init.normal_(p, 0, 0.1)
        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'] = automap
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
MiniMaxM2ForCausalLM(
  (model): MiniMaxM2Model(
    (embed_tokens): Embedding(200064, 8)
    (layers): ModuleList(
      (0-1): 2 x MiniMaxM2DecoderLayer(
        (self_attn): MiniMaxM2Attention(
          (q_proj): Linear(in_features=8, out_features=256, bias=False)
          (k_proj): Linear(in_features=8, out_features=128, bias=False)
          (v_proj): Linear(in_features=8, out_features=128, bias=False)
          (o_proj): Linear(in_features=256, out_features=8, bias=False)
          (q_norm): MiniMaxM2RMSNorm((256,), eps=1e-06)
          (k_norm): MiniMaxM2RMSNorm((128,), eps=1e-06)
        )
        (block_sparse_moe): MiniMaxM2SparseMoeBlock(
          (gate): Linear(in_features=8, out_features=32, bias=False)
          (experts): MiniMaxM2Experts(
            (0-31): 32 x MiniMaxM2MLP(
              (w1): Linear(in_features=8, out_features=64, bias=False)
              (w2): Linear(in_features=64, out_features=8, bias=False)
              (w3): Linear(in_features=8, out_features=64, bias=False)
              (act_fn): SiLUActivation()
            )
          )
        )
        (input_layernorm): MiniMaxM2RMSNorm((8,), eps=1e-06)
        (post_attention_layernorm): MiniMaxM2RMSNorm((8,), eps=1e-06)
      )
    )
    (norm): MiniMaxM2RMSNorm((8,), eps=1e-06)
    (rotary_emb): MiniMaxM2RotaryEmbedding()
  )
  (lm_head): Linear(in_features=8, out_features=200064, bias=False)
)
```