Gluon: a GPU programming language based on the same compiler stack as Triton (github.com)

Gluon is a new, lower-level GPU programming language built on the same Triton compiler and Python DSL/JIT stack, announced as an experimental extension in the Triton repository. Unlike Triton — which hides many hardware and layout choices behind compiler heuristics — Gluon exposes those low-level details (tile/layout selection, memory allocation, data movement, and asynchrony) so developers can hand-control CTA/grid mapping and optimize kernels more finely. That tradeoff: more programmer responsibility, but the potential to out‑perform Triton’s general-purpose code for performance-critical kernels. Technically, Gluon keeps Triton’s tile-based SPMD model and the same host/launcher semantics (PyTorch tensors become global pointers; kernels are declared with @gluon.jit and launched with a grid and num_warps). It supports constexpr hyperparameters and integrates with Triton’s autotuner, demonstrated with memcpy kernels that tune block sizes (XBLOCK) and reported GB200 results (best XBLOCK=2048, ~666 GB/s for an 8 GB copy). Tutorials in the repo walk from basic loads/stores to advanced features — layouts, async copy/TMA, warp specialization, persistence and GEMM — making Gluon useful for developers willing to trade abstraction for deterministic, hand-tuned performance on modern GPUs.