Apache TVM

Apache TVM: The open-source deep learning compiler framework for universal model optimization and deployment.

Apache TVM is the machine learning compilation stack that bridges the gap between high-level frameworks (PyTorch, ONNX) and diverse hardware. It ingests pre-trained models, then executes a series of composable optimizations (graph and tensor-level) to generate highly efficient, deployable modules. TVM is hardware-agnostic: it targets everything from standard CPUs and GPUs (x86, CUDA) to specialized accelerators and embedded systems (ARM, TPUs, Metal, Vulkan). The goal is clear: maximize model performance, reduce latency, and ensure Python-first development leads to universal, high-speed execution across any target device.

https://tvm.apache.org/