PyTorch-Triton 3.7 introduces Plugin Extensions for custom compiler passes

The PyTorch-Triton 3.7 release introduces Triton Plugin Extensions, a general-purpose system that dynamically loads custom compiler passes, MLIR dialects, and DSL extensions into upstream Triton at runtime—without forking or recompiling. Until now, advanced GPU kernel optimizations (custom passes, hardware intrinsics, specialized memory patterns) forced teams to maintain forks of Triton, which quickly fell behind upstream and created merge conflicts. The new plugin architecture solves this by exposing hooks at every lowering level—TTIR, TTGIR, LLVM IR, and target-specific assembly (PTX, AMDGCN). Plugins are shared libraries (.so files) discovered via the TRITON_PLUGIN_PATHS environment variable. They can insert, disable, or replace passes at arbitrary points in the pipeline, and even define custom MLIR dialects. Per-kernel control allows toggling custom pipelines dynamically, with plugins managing their own cache invalidation.

As the first major consumer, Meta's Triton Language Extensions (TLX) are now enabled out of the box. TLX brings persistent GEMM kernels and fine-grained hardware control to stock Triton, with performance matching or exceeding vendor libraries on both NVIDIA H100 and AMD MI350. Developers can activate TLX simply by setting an environment variable pointing to the libutlx.so shared library. The plugin system also supports custom top-level DSL ops, enabling entirely new programming abstractions without altering Triton itself. This open architecture allows researchers and engineers to iterate on custom features at full speed while always running on the latest upstream release, eliminating the maintenance burden of forked repos. Triton Plugin Extensions are available for both NVIDIA and AMD backends, paving the way for a vibrant ecosystem of community-contributed compiler optimizations.