---
id: nvidia-cutlass-dsl-libs-core
version: "4.7.0"
license: unclear
license_treatment: unclear
maintenance: active
---

# nvidia-cutlass-dsl-libs-core — NVIDIA CUTLASS Python DSL

License: unclear · Maintenance: active · Downloads: 1.1M/mo

## What it is and what it does

nvidia-cutlass-dsl-libs-core is a Python wrapper around NVIDIA's CUTLASS 4.x, exposing CuTe DSL—a domain-specific language for GPU kernel programming. It lets you write high-performance CUDA kernels in Python without C++ expertise, targeting Tensor Cores on modern NVIDIA GPUs. The package handles layouts, tensors, hardware atoms, and thread/data hierarchy control, with a focus on matrix multiply and linear algebra operations.

The package is currently in public beta and aims to flatten the GPU programming learning curve. It depends on numpy, protobuf, cuda-python, and nvidia-cuda-nvdisasm, so your environment must have CUDA 13 and an NVIDIA GPU. Compile times are claimed to be orders of magnitude faster than C++ alternatives, and it integrates natively with deep learning frameworks.

Use it for:

- Prototype optimized matrix multiply kernels for Tensor Cores without writing C++
- Rapidly iterate on custom GPU kernels for deep learning research and experimentation
- Integrate hand-tuned CUDA kernels into deep learning workflows without C++ bindings
- Learn GPU programming and CuTe abstractions with a lower barrier to entry than C++
- Deploy production-grade linear algebra kernels targeting Ampere, Hopper, or Blackwell GPUs

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Provides a Python interface for writing high-performance CUDA kernels using CuTe DSL abstractions, targeting NVIDIA Tensor Cores on Ampere, Hopper, and Blackwell architectures without requiring deep C++ expertise.

Yes, if you have a CUDA 13 environment and need to write or prototype GPU kernels without C++ expertise. The low install friction, active maintenance, and recent release are positive signals. However, the public beta status and unclear license terms warrant caution for production use—verify licensing and test stability for your workload before committing.

## Install

pip install nvidia-cutlass-dsl-libs-core
uv add nvidia-cutlass-dsl-libs-core
poetry add nvidia-cutlass-dsl-libs-core

## Installing nvidia-cutlass-dsl-libs-core

Before you install: Low install friction with a pure-wheel distribution. Active maintenance—released 9 days ago with recent commits. Depends on cuda-python and nvidia-cuda-nvdisasm, which require NVIDIA GPU tooling; verify your CUDA environment before installing.

License in practice: License treatment is unclear; no SPDX identifier or raw license text is available. Classified as proprietary. Verify licensing terms before using in production or redistribution.

Quickstart:

pip install nvidia-cutlass-dsl-libs-core
import nvidia_cutlass_dsl_libs_core
# Use CuTe DSL abstractions for kernel design

Requires CUDA 13 environment and NVIDIA GPU; Python >=3.10; cuda-python and nvidia-cuda-nvdisasm must be available.

Verify before relying:

- Exact scope of CuTe DSL API surface and whether it covers all linear algebra operations mentioned
- Performance benchmarks vs. native C++ CUTLASS on the same architectures
- Stability guarantees given the public beta status and planned graduation by summer 2026
- Whether deep learning framework integration is automatic or requires additional glue code

## Package facts

- License: not declared (unclear)
- Python support: supports_current
- Install friction: low
- Maintenance: active
- Downloads: 1.1M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

cuda kernel programming python, gpu tensor core optimization, cute dsl python interface, nvidia cutlass python bindings, high-performance cuda kernels, matrix multiply gpu acceleration, ampere hopper blackwell kernels, gpu-programming, cuda-kernels, tensor-cores

[View on SkillFed](https://skillfed.io/packages/nvidia-cutlass-dsl-libs-core) · [View on PyPI](https://pypi.org/project/nvidia-cutlass-dsl-libs-core/)