Deplodock compiles LLMs to raw CUDA

The point here is not to beat Triton everywhere; it is to make compiler structure legible enough that you can hack on it without spelunking through TVM-sized codebases.

• –Six IRs give a clean mental model from FX capture to CUDA emission, which is more useful for learning and experimentation than a monolithic lowering pass.
• –The fusion and staging story is the real value: for LLM inference, avoiding intermediate HBM traffic usually matters more than clever algebra.
• –Raw CUDA output makes every scheduling choice inspectable, but it also means portability and peak-kernel performance will lag production-grade stacks.
• –The RMSNorm, softmax, and matmul examples make clear this is a compiler tutorial first and a benchmark race second.
• –For AI developers, the project is interesting as a readable reference architecture, not as a drop-in replacement for torch.compile.