Forge CLI generates GPU kernels optimized for any PyTorch or HuggingFace model. It runs a swarm of 32 parallel Coder and Judge agents that compete to produce the fastest CUDA/Triton implementation, achieving up to five times the speed of torch.compile(mode='max‑autotune') while maintaining 97.6 % correctness. Users provide a HuggingFace model ID and receive optimized kernels for each layer, improving inference speed and GPU utilization with minimal effort.