File size: 5,632 Bytes
2f0a56e 809058a 2f0a56e 809058a 2f0a56e |
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 |
---
library_name: transformers
pipeline_tag: text-generation
inference: true
widget:
- text: Hello!
example_title: Hello world
group: Python
---
This tiny model is for debugging. It is randomly initialized with the config adapted from [MiniMaxAI/MiniMax-M1-80k](https://huggingface.co/MiniMaxAI/MiniMax-M1-80k).
### Example usage:
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
model_id = "tiny-random/minimax-m1"
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.'))
```
### Printing the model:
```text
MiniMaxM1ForCausalLM(
(model): MiniMaxM1Model(
(embed_tokens): Embedding(200064, 64)
(layers): ModuleList(
(0): MiniMaxM1DecoderLayer(
(self_attn): MiniMaxM1LightningAttention(
(out_proj): Linear(in_features=64, out_features=64, bias=False)
(norm): MiniMaxM1RMSNorm()
(qkv_proj): Linear(in_features=64, out_features=192, bias=False)
(output_gate): Linear(in_features=64, out_features=64, bias=False)
)
(block_sparse_moe): MiniMaxM1SparseMoeBlock(
(gate): Linear(in_features=64, out_features=8, bias=False)
(experts): ModuleList(
(0-7): 8 x MiniMaxM1BlockSparseTop2MLP(
(w1): Linear(in_features=64, out_features=128, bias=False)
(w2): Linear(in_features=128, out_features=64, bias=False)
(w3): Linear(in_features=64, out_features=128, bias=False)
(act_fn): SiLU()
)
)
)
(input_layernorm): MiniMaxM1RMSNorm()
(post_attention_layernorm): MiniMaxM1RMSNorm()
)
(1): MiniMaxM1DecoderLayer(
(self_attn): MiniMaxM1FlashAttention2(
(q_proj): Linear(in_features=64, out_features=64, bias=False)
(k_proj): Linear(in_features=64, out_features=32, bias=False)
(v_proj): Linear(in_features=64, out_features=32, bias=False)
(o_proj): Linear(in_features=64, out_features=64, bias=False)
(rotary_emb): MiniMaxM1RotaryEmbedding()
)
(block_sparse_moe): MiniMaxM1SparseMoeBlock(
(gate): Linear(in_features=64, out_features=8, bias=False)
(experts): ModuleList(
(0-7): 8 x MiniMaxM1BlockSparseTop2MLP(
(w1): Linear(in_features=64, out_features=128, bias=False)
(w2): Linear(in_features=128, out_features=64, bias=False)
(w3): Linear(in_features=64, out_features=128, bias=False)
(act_fn): SiLU()
)
)
)
(input_layernorm): MiniMaxM1RMSNorm()
(post_attention_layernorm): MiniMaxM1RMSNorm()
)
)
(norm): MiniMaxM1RMSNorm()
)
(lm_head): Linear(in_features=64, out_features=200064, bias=False)
)
```
### Codes to create this repo:
```python
import json
from pathlib import Path
import torch
import accelerate
from huggingface_hub import file_exists, hf_hub_download
from transformers import (
AutoConfig,
AutoModelForCausalLM,
AutoTokenizer,
GenerationConfig,
set_seed,
)
source_model_id = "MiniMaxAI/MiniMax-M1-80k"
save_folder = "/tmp/tiny-random/minimax-m1"
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"] = [0, 1] # one lightning, one attention
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'] = 64
config_json['intermediate_size'] = 128
config_json['num_attention_heads'] = 2
config_json['num_experts_per_tok'] = 2
config_json['num_hidden_layers'] = 2
config_json['num_key_value_heads'] = 1
config_json['num_local_experts'] = 8
config_json['rotary_dim'] = 16
config_json['tie_word_embeddings'] = True
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() # 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'] = 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()
``` |