#safetensors #tensor #ndarray #serialization #array

ndarray-safetensors

Serialize / deserialize ndarrays with safetensors

5 unstable releases

0.3.0 Mar 5, 2025
0.2.2 Jan 9, 2025
0.2.1 Jan 5, 2025
0.2.0 Dec 22, 2024
0.1.0 Dec 21, 2024

#542 in Encoding

Download history 243/week @ 2024-12-18 16/week @ 2024-12-25 166/week @ 2025-01-01 159/week @ 2025-01-08 3/week @ 2025-02-05 14/week @ 2025-02-12 2/week @ 2025-02-19 23/week @ 2025-02-26 141/week @ 2025-03-05 24/week @ 2025-03-12

190 downloads per month
Used in e2k

MIT license

37KB
663 lines

ndarray-safetensors

Serialize / deserialize Rust ndarray with Safetensors.

Demo

See safetensors-testapp for more details. Here is an end-to-end demo to print all tensors from a safetensors file.

use ndarray_safetensors::parse_tensor_view_data;

let path_example = Path::new("data/rand.safetensors");
let file = File::open(path_example).unwrap();
let buffer = unsafe {MmapOptions::new().map(&file).unwrap()};
let tensors = safetensors::SafeTensors::deserialize(&buffer).unwrap();
for (name, tensor_view) in tensors.tensors() {
    println!("Tensor: {}", name);
    if tensor_view.dtype() == safetensors::Dtype::F32 {
        let arr= parse_tensor_view_data::<f32>(&tensor_view).unwrap();
        println!("{}", arr);
    }
    if tensor_view.dtype() == safetensors::Dtype::F64 {
        let arr= parse_tensor_view_data::<f64>(&tensor_view).unwrap();
        println!("{}", arr);
    }
}

Features

The following crate feature flags are available. They are configured in your Cargo.toml.

  • unsafe_copy: Allowing directly copying the bytes instead of parsing floats in serialization/deserialization for little-endian machines.
  • x86_sse: Using SSE instructions to optimize some conversions, it could be useful for fp16.

Copyright (c) 2024-2025 Mengxiao Lin. Check LICENSE file.

Dependencies

~2–3MB
~64K SLoC