microsoft/nnfusion

Health v2 · maintenance only

Adoption not part of the score

1002 stars · 166 forks observed · 2026-09-03

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

NNFusion is a flexible and efficient deep neural network (DNN) compiler that generates high-performance executables from model descriptions such as TensorFlow frozen models and ONNX files. It performs full-stack optimizations including kernel fusion, memory layout optimization, and ahead-of-time code generation with minimal runtime dependencies.

Use cases

• compile a frozen TensorFlow or ONNX model into a fast standalone executable
• deploy a pre-trained DNN model without framework dependencies
• speed up inference of a pre-defined deep learning model on CUDA or ROCm GPUs
• experiment with custom compiler optimizations for specific neural network models
• generate human-readable source code from a model to hand-tune kernels
• run distributed parallel training with SuperScaler integration

When to choose

• you need framework-free, low-overhead deployment of TensorFlow or ONNX models on GPUs
• you want ahead-of-time compilation instead of runtime graph execution
• you are a researcher prototyping compiler optimizations for DNNs
• you need kernel fusion and memory layout tuning for inference performance

When to avoid

• you need broad model format support beyond TensorFlow and ONNX
• you primarily target Windows or macOS, since support is focused on Ubuntu with CUDA
• you want an actively evolving project, as development appears to be in maintenance mode
• you need dynamic-shape or training-focused compilation without extra tooling

Facets

cli-tool · maturity maintenance

compiler machine-learning deep-learning llm-inference gpu-computing developer-tools machine-learning deep-learning compilers gpu-computing developer-tools cpp dnn-compiler code-generation onnx tensorflow kernel-fusion ahead-of-time-compilation inference linux docker gpu

Member repositories

For agents

markdown · JSON · MCP: product_card(name="microsoft/nnfusion")