TVM is an automated compiler designed to optimize deep learning workloads for a wide range of hardware back-ends, including CPUs, GPUs, and specialized accelerators. It addresses the challenges of deploying deep learning models on diverse hardware platforms by providing both graph-level and operator-level optimizations. TVM uses a tensor expression language to describe operators and a machine learning-based cost model to guide the optimization process. The system is open-sourced and has been evaluated on various platforms, demonstrating competitive performance with state-of-the-art hand-tuned libraries. TVM's ability to handle new hardware back-ends and emerging workloads, such as depthwise convolution and low-precision operations, is highlighted through experimental results.TVM is an automated compiler designed to optimize deep learning workloads for a wide range of hardware back-ends, including CPUs, GPUs, and specialized accelerators. It addresses the challenges of deploying deep learning models on diverse hardware platforms by providing both graph-level and operator-level optimizations. TVM uses a tensor expression language to describe operators and a machine learning-based cost model to guide the optimization process. The system is open-sourced and has been evaluated on various platforms, demonstrating competitive performance with state-of-the-art hand-tuned libraries. TVM's ability to handle new hardware back-ends and emerging workloads, such as depthwise convolution and low-precision operations, is highlighted through experimental results.