Safetensors

Safetensors is a new simple format for storing tensors safely (as opposed to pickle) and that is still fast (zero-copy). Safetensors is really fast 🚀.

Installation

with pip:

pip install safetensors

with conda:

conda install -c huggingface safetensors

Usage

Load tensors

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)

Loading only part of the tensors (interesting when running on multiple GPU)

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

import torch
from safetensors.torch import save_file

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

Format

Let’s say you have safetensors file named model.safetensors, then model.safetensors will have the following internal format:

Featured Projects

Safetensors is being used widely at leading AI enterprises, such as Hugging Face, EleutherAI, and StabilityAI. Here is a non-exhaustive list of projects that are using safetensors:

< > Update on GitHub