The authors replace the conventional Triton-to-PTX compiler backend with an LLM agent that generates PTX directly from Triton source. They build an environment that evaluates candidate PTX for correctness and performance, plus an agentic harness that iteratively revises the generated code. This approach, called AI lowering, bypasses hand-written intermediate representations and lowering passes entirely.
Across twelve common kernels on Ada, Hopper, and Blackwell GPUs and ten kernels from recent ML papers, AI lowering achieves 0.83x-3.34x the performance of autotuned Triton. The biggest wins come from transformations Triton's pipeline does not perform, such as decoding packed binary weights into Tensor Core operands (3.34x on BitDelta), assigning each thread a complete softmax row in tensor memory (1.37x on FlashAttention), and reusing overlapping convolution windows (up to 2.23x). The work also extends the Volta PTX verifier to support Blackwell's tcgen05 interface, including managed tensor memory, descriptor-based operand layouts, and asynchronous execution.