File size: 1,607 Bytes
0bf108f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
library_name: transformers
pipeline_tag: text-generation
inference: true
widget:
- text: Hello!
  example_title: Hello world
  group: Python
---

This model is for debugging. It is randomly initialized using the config from [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) but with smaller size. 

Codes:
```python
import os

import torch
import transformers
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, GenerationConfig, pipeline, set_seed

model_id = "meta-llama/Llama-3.2-3B-Instruct"
repo_id = "yujiepan/llama-3.2-tiny-random"
save_path = f"/tmp/{repo_id}"

config = AutoConfig.from_pretrained(model_id, trust_remote_code=True)
config._name_or_path = model_id
config.hidden_size = 8
config.intermediate_size = 16
config.num_attention_heads = 2
config.num_key_value_heads = 1
config.head_dim = 4
config.num_hidden_layers = 2
config.torch_dtype = "bfloat16"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
tokenizer.save_pretrained(save_path)

model = AutoModelForCausalLM.from_config(
    config, torch_dtype=torch.bfloat16, attn_implementation="sdpa", trust_remote_code=True
)
model.generation_config = GenerationConfig.from_pretrained(
    model_id, trust_remote_code=True)

set_seed(42)
with torch.no_grad():
    for _, p in sorted(model.named_parameters()):
        torch.nn.init.uniform_(p, -0.2, 0.2)

model.save_pretrained(save_path)

pipe = pipeline("text-generation", model=save_path, device="cpu",
                trust_remote_code=True, max_new_tokens=20)
print(pipe("Hello World!"))
```