dlpack-rs

Portable Rust bindings for the DLPack tensor protocol.

• dlpack_rs::ffi — hand-written #[repr(C)] ABI types (DLTensor, DLManagedTensor, DLManagedTensorVersioned, DLDevice, DLDataType, DLPackVersion). No bindgen, no C submodule, no build script.
• dlpack_rs::safe — ownership-correct pack / pack_versioned builders that wrap a Rust keepalive + shape/strides Vecs in a heap-allocated DLManagedTensor[Versioned] with a correct deleter callback.
• dlpack_rs::pyo3_glue (feature pyo3) — IntoDLPack export trait + PyTensor import struct; implements the __dlpack__ capsule lifecycle (rename-on-consume + deleter-on-drop).

Installation

Add to Cargo.toml:

# no pyo3 glue (pure Rust, no Python dependency):
dlpack-rs = { git = "https://github.com/kornia/dlpack-rs", tag = "v0.2.0" }

# with pyo3 glue:
dlpack-rs = { git = "https://github.com/kornia/dlpack-rs", tag = "v0.2.0", features = ["pyo3"] }

Usage

Raw FFI

use dlpack_rs::ffi::{DLDataType, DLDevice, DLTensor, K_DL_CPU, K_DL_FLOAT};

let device = DLDevice { device_type: K_DL_CPU, device_id: 0 };
let dtype  = DLDataType { code: K_DL_FLOAT, bits: 32, lanes: 1 };

Safe builder (safe::pack)

use dlpack_rs::safe::{pack, cpu_device, dtype_f32, TensorInfo};

let mut data = vec![1.0f32, 2.0, 3.0, 4.0];
let info = TensorInfo::contiguous(
    data.as_mut_ptr() as *mut _,
    cpu_device(),
    dtype_f32(),
    vec![2, 2],
);

// `data` is the keepalive: it will be dropped when the DLPack deleter fires.
let mt: *mut dlpack_rs::ffi::DLManagedTensor = pack(data, info);

// Hand `mt` to a consumer (e.g. Python via a capsule).
// The consumer calls mt.deleter(mt) when done; that drops the keepalive.

pyo3 __dlpack__ (feature pyo3)

Export — give a Python caller a zero-copy view of Rust data:

use dlpack_rs::pyo3_glue::IntoDLPack;
use dlpack_rs::safe::{cpu_device, dtype_f32, TensorInfo};
use pyo3::prelude::*;
use std::ffi::c_void;

struct MyTensor(Vec<f32>);

impl IntoDLPack for MyTensor {
    fn tensor_info(&self) -> TensorInfo {
        TensorInfo::contiguous(
            self.0.as_ptr() as *mut c_void,
            cpu_device(),
            dtype_f32(),
            vec![self.0.len() as i64],
        )
    }
}

Python::attach(|py| {
    let t = MyTensor(vec![1.0, 2.0, 3.0]);
    let capsule = t.into_capsule(py).unwrap();
    // pass `capsule` to Python as the return value of `__dlpack__()`
});

Import — consume a capsule from any Python object that implements __dlpack__():

use dlpack_rs::pyo3_glue::PyTensor;
use pyo3::prelude::*;

Python::attach(|py| {
    let obj: Bound<'_, PyAny> = /* ... numpy array, torch tensor, etc. */ todo!();
    let tensor = PyTensor::from_pyany(py, &obj).unwrap();

    println!("shape  = {:?}", tensor.shape());
    println!("dtype  = {:?}", tensor.dtype());
    println!("device = {:?}", tensor.device());
    println!("nbytes = {}", tensor.nbytes());

    // When `tensor` is dropped, the producer's DLPack deleter is called exactly once.
});

Feature flags

Feature	Default	Description
pyo3	no	Enables pyo3_glue (requires Python headers at build time)

Without any features (--no-default-features), the crate is a pure no_std-compatible FFI + safe-builder layer with zero non-std dependencies.

License

Apache-2.0 — see LICENSE.