@vikhyatk Describes Photon 2.0 Megakernel Compiler

Vikhyat K., co-founder and CTO of Moondream AI, posted that the team grew tired of hand-tuning GPU kernels and built Photon 2.0 instead. The compiler converts the complete forward pass of Moondream, Qwen 3.5, and Gemma 4 into megakernels that run as single GPU programs. He illustrated the approach with a tracer DSL example that maps data flows and dependencies for a dense Moondream layer. Other posts in the thread include images showing resulting GPU behavior.