AI as a Compiler: Compiling Triton kernels without the Triton compiler

An LLM agent writes GPU assembly straight from Triton code, sometimes beating Triton's own compiler.

3.34×

speedup

Best speedup over autotuned Triton

Gained by decoding packed binary weights directly into Tensor Core operands

How it works

Compilers are costly
Every new GPU feature needs hand-written lowering passes across several intermediate representations, which are slow to build and maintain.
LLM writes PTX directly
An agent generates PTX from Triton source. It is tested, profiled, and patched in loops until the budget runs out.
Parity to 3.34× faster
Generated kernels range from 0.83× to 3.34× of Triton speed. The biggest wins come from tricks Triton never attempts.

From the paper