Show HN: Compile computation graphs into transformer weights – no training (github.com)

The recent launch of **torchwright**, a compiler that converts computation graphs into transformer weights without the need for training, marks a significant advancement in the AI/ML community. Users simply define a computation graph using standard Python, and torchwright generates the weights for a transformer model, specifically targeting the stock Phi-3 architecture. This eliminates the traditional training phase, allowing for the rapid deployment of customized transformer models in scenarios where training data may be limited or impractical. Key technical characteristics of torchwright include its ability to compile models in under ten seconds on a standard laptop CPU, using a Directed Acyclic Graph (DAG) structure for operations. The compiler utilizes various types of operations divided into linear, attention, and nonlinear categories, ensuring that every operation is directly translated into concrete weights utilized in attention heads or feedforward neural network sublayers. Furthermore, the built-in optimization capabilities allow users to employ constraint programming to minimize layer counts, contrasting with existing languages like RASP that offer less flexibility. With its capability to produce Hugging Face-compatible checkpoints, torchwright not only streamlines the model creation process but also fosters innovation in how AI models can be adapted and deployed across diverse applications.