PyTorch-Triton 3.7 introduces a plugin extensions system that lets developers dynamically load custom compiler passes, dialects, and operations into upstream Triton without forking or recompiling it. Meta’s Triton Language Extensions (TLX), which provide persistent GEMM kernels and fine-grained hardware control through operations like tlx.local_alloc(), tlx.async_load(), and tlx.async_dot(), are now available as a standalone Python package that works with unmodified upstream Triton. Performance validation shows TLX matching or exceeding vendor libraries on Nvidia H100 (on par with cuBLAS on square GEMMs, faster on wide shapes) and delivering 12-15% higher throughput than rocBLAS on AMD MI350, with zero overhead from dynamic loading.
