File size: 5,925 Bytes
e9a82a2 |
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 |
---
library_name: transformers
base_model:
- MiniMaxAI/MiniMax-M2.5
---
This tiny model is intended for debugging. It is randomly initialized using the configuration adapted from [MiniMaxAI/MiniMax-M2.5](https://huggingface.co/MiniMaxAI/MiniMax-M2.5).
| File path | Size |
|------|------|
| model.safetensors | 7.1MB |
### Example usage:
- vLLM
```bash
vllm serve tiny-random/minimax-m2.5 --trust-remote-code --reasoning-parser minimax_m2_append_think --enable-auto-tool-choice --tool-call-parser minimax_m2
```
- Transformers
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
model_id = "tiny-random/minimax-m2.5"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
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.', max_new_tokens=16))
```
### Codes to create this repo:
<details>
<summary>Click to expand</summary>
```python
import json
from pathlib import Path
import accelerate
import torch
import transformers
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
GenerationConfig,
set_seed,
)
# try:
# from transformers.utils.output_capturing import OutputRecorder, capture_outputs
# transformers.utils.generic.OutputRecorder = OutputRecorder
# transformers.utils.generic.capture_outputs = capture_outputs
# transformers.utils.generic.check_model_inputs = capture_outputs
# transformers.modeling_rope_utils.ROPE_INIT_FUNCTIONS['default'] = transformers.modeling_rope_utils.ROPE_INIT_FUNCTIONS['linear']
# except ImportError:
# pass
source_model_id = "MiniMaxAI/MiniMax-M2.5"
save_folder = "/tmp/tiny-random/minimax-m25"
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]
# del config_json['auto_map']
# del config_json['num_mtp_modules']
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'] = 32
config_json['num_attention_heads'] = 8
config_json['num_key_value_heads'] = 4
config_json['num_hidden_layers'] = 2
config_json['mlp_intermediate_size'] = 32
# 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,
)
# config.standardize_rope_params()
# config.rope_parameters['rope_type'] = 'linear'
# config.rope_parameters['factor'] = 1.0
torch.set_default_dtype(torch.bfloat16)
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
torch.set_default_dtype(torch.float32)
print(model)
# according to source model, gate is in FP32
for i in range(config.num_hidden_layers):
model.model.layers[i].block_sparse_moe.gate = model.model.layers[i].block_sparse_moe.gate.float()
model.model.layers[i].block_sparse_moe.e_score_correction_bias = model.model.layers[i].block_sparse_moe.e_score_correction_bias.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)
automap = config_json['auto_map']
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()
```
</details>
### Printing the model:
<details><summary>Click to expand</summary>
```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=256, bias=False)
(experts): MiniMaxM2Experts(
(0-255): 256 x MiniMaxM2MLP(
(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()
)
)
)
(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)
)
```
</details> |