# Safetensors

## Huggingface Safetensors

Safetensors는 텐서를 안전하게 저장하는(ML에서 자주 사용하는 pikle과 반대되는) 새로운 간단한 형식이며, 장점은 Zerocopy로 빠릅니다.

![saftensors](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/safetensors/safetensors-format.svg)

실제 Huggingface에서 모델을 push 하면 자동적으로 safetensors로 변환되기 때문에 대다수 Model이 safetensors로 이루어져 있다.

<figure><img src="https://3055094660-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FYzxz4QeW9UTrhrpWwKiQ%2Fuploads%2FTzihLCwLzYOVQ6WaeukL%2F%E1%84%89%E1%85%B3%E1%84%8F%E1%85%B3%E1%84%85%E1%85%B5%E1%86%AB%E1%84%89%E1%85%A3%E1%86%BA%202024-05-23%20%E1%84%8B%E1%85%A9%E1%84%92%E1%85%AE%2011.41.12.png?alt=media&#x26;token=a6c827e7-65be-4ce6-a8cc-b9c08d3b288a" alt=""><figcaption></figcaption></figure>

```python
%pip install safetensors
```

```
Requirement already satisfied: safetensors in /home/kubwa/anaconda3/envs/pytorch/lib/python3.11/site-packages (0.4.2)
Note: you may need to restart the kernel to use updated packages.
```

### Load tensors

```python
from safetensors import safe_open

tensors = {}
with safe_open("model.safetensors", framework="pt", device=0) as f:
    for k in f.keys():
        tensors[k] = f.get_tensor(k)
```

```python
from safetensors import safe_open

tensors = {}
with safe_open("model.safetensors", framework="pt", device=0) as f:
    tensor_slice = f.get_slice("embedding")
    vocab_size, hidden_dim = tensor_slice.get_shape()
    tensor = tensor_slice[:, :hidden_dim]
```

### Save tensors

```python
import torch
from safetensors.torch import save_file

tensors = {
    "embedding": torch.zeros((2, 2)),
    "attention": torch.zeros((2, 3))
}
save_file(tensors, "model.safetensors")
```