dmlc/dlpack

Health v2 · maintenance only

Adoption not part of the score

1242 stars · 168 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

DLPack is a minimal, stable open in-memory tensor structure defined as a single C ABI header, enabling zero-copy tensor exchange between deep learning frameworks and array libraries. It is a widely adopted standard supported by NumPy, PyTorch, TensorFlow, JAX, CuPy, TVM, and others across CPU, CUDA, ROCm, Vulkan, Metal, and WebGPU devices.

Use cases

• share tensors between pytorch and tensorflow without copying
• wrap vendor gpu operator libraries in a framework-agnostic way
• exchange ndarrays between numpy and cupy across devices
• build a compiler backend that consumes tensors from any framework
• swap blas or tensor backend implementations at runtime
• expose a c-compatible tensor interface from another language

When to choose

• you need zero-copy tensor interchange between frameworks or languages
• you target multiple hardware backends (cuda, rocm, vulkan, metal, webgpu)
• you want a minimal, stable, allocator-free C ABI standard with wide adoption

When to avoid

• you need a full tensor library with ops and allocators - dlpack deliberately provides neither
• you need rich layout features beyond row-major with strides, or reference-counted ownership semantics beyond its scope

Facets

library · maturity stable

serialization sdk deep-learning machine-learning gpu-computing developer-tools cpp cross-platform python tensor-interoperability c-abi header-only zero-copy array-exchange gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="dmlc/dlpack")