---
id: nvidia-cudnn-frontend
version: "1.27.0"
license: Apache-2.0 AND MIT
license_treatment: permissive
maintenance: active
---

# nvidia-cudnn-frontend — NVIDIA cuDNN Frontend — Python and C++ Graph API with SOTA attention (SDPA / Flash Attention), MoE grouped GEMM fusions, and FP8/MXFP8 kernels for Hopper and Blackwell GPUs.

License: permissive · Maintenance: active · Downloads: 4.5M/mo

## What it is and what it does

nvidia-cudnn-frontend is NVIDIA's modern entry point to cuDNN, offering both a header-only C++ API and Python bindings that abstract the complexity of the cuDNN Graph API. It exposes state-of-the-art GPU kernels including scaled dot-product attention (SDPA/Flash Attention), grouped GEMM fusions for mixture-of-experts training, fused normalization and activation operations, and quantized matrix multiplication in FP8 and MXFP8 precision. The package targets NVIDIA's latest GPU architectures—Hopper (H100/H200) and Blackwell (B200/GB200/GB300)—and includes native PyTorch integration with torch.compile support.

The package ships with open-source kernel implementations (FROST GEMM engine, block-sparse attention, native sparse attention, and fused RMSNorm+SiLU) that developers can inspect, modify, and contribute to. Installation is simple via pip, though it requires NVIDIA driver, CUDA Toolkit, and cuDNN 8.5.0 or later on the system. The library is actively maintained with prebuilt wheels for modern Python versions and multiple architectures.

Use it for:

- Accelerate transformer attention mechanisms in large language models using SDPA kernels optimized for Hopper and Blackwell GPUs.
- Train mixture-of-experts models efficiently with fused grouped GEMM operations that reduce memory bandwidth and kernel launch overhead.
- Deploy quantized deep learning models using FP8 and MXFP8 precision kernels for reduced memory footprint and faster inference.
- Build custom GPU kernels by inspecting and modifying open-source implementations like FROST GEMM and block-sparse attention.
- Integrate cuDNN-accelerated operations into PyTorch models with automatic differentiation and torch.compile support.

## Worth the install?

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

Provides Python and C++ APIs to NVIDIA's cuDNN library, exposing high-performance GPU kernels for scaled dot-product attention, grouped matrix multiplication for mixture-of-experts training, and fused operations optimized for Hopper and Blackwell GPUs.

Yes. The package is actively maintained, dual-licensed under permissive terms, has no known vulnerabilities, and provides essential optimized kernels for modern GPU workloads. Install it if you are training or deploying deep learning models on NVIDIA Hopper or Blackwell GPUs and need high-performance attention, grouped GEMM, or quantized operations. The requirement for CUDA Toolkit and cuDNN 8.5.0+ is a prerequisite, not a drawback—it reflects the package's tight integration with NVIDIA's GPU stack.

## Install

pip install nvidia-cudnn-frontend
uv add nvidia-cudnn-frontend
poetry add nvidia-cudnn-frontend

## Installing nvidia-cudnn-frontend

Before you install: Installation is straightforward via pip with prebuilt wheels for Python 3.10–3.14 on Linux (x86_64 and aarch64) and Windows. The package is actively maintained with a recent release (8 days old) and no known vulnerabilities. Medium install friction reflects the requirement for NVIDIA driver, CUDA Toolkit, and cuDNN 8.5.0 or later to be present on the system.

License in practice: Dual-licensed under Apache-2.0 and MIT, both permissive licenses. You are free to use, modify, and distribute the package in commercial and private projects with minimal restrictions.

Quickstart:

pip install nvidia-cudnn-frontend

import cudnn

# Create a graph for scaled dot-product attention
graph = cudnn.pygraph.Graph()
# Configure and execute attention operations on Hopper/Blackwell GPUs

Requires NVIDIA driver, CUDA Toolkit, and cuDNN 8.5.0 or later installed on the system; Python 3.9 or later.

Verify before relying:

- Whether the package's autotuning mechanism requires additional configuration or warm-up time for production workloads.
- Performance characteristics and memory overhead of the open-source FROST GEMM engine relative to backend-native plans.
- Compatibility and integration effort with existing PyTorch models beyond the stated torch.compile support.

## Package facts

- License: Apache-2.0 AND MIT (permissive)
- Python support: supports_current
- Install friction: medium
- Maintenance: active
- Downloads: 4.5M/month (top 5,000 on PyPI)
- Known vulnerabilities: none known

## Tags

cudnn python bindings, gpu attention kernels, flash attention implementation, mixture of experts gemm, nvidia cuda deep learning, fp8 matrix multiplication, hopper blackwell gpu kernels, gpu-acceleration, transformer-kernels, quantization

[View on SkillFed](https://skillfed.io/packages/nvidia-cudnn-frontend) · [View on PyPI](https://pypi.org/project/nvidia-cudnn-frontend/)