Triton hits PyTorch 2.11 with Blackwell support

Triton has effectively become the industry standard for breaking the "CUDA moat" by bringing high-performance kernel development into the Python ecosystem.

• –FlashAttention-4 backend delivers 1.2x to 3.2x speedups over previous Triton implementations for compute-bound Blackwell workloads.
• –Native support for FP8 and FP16 GEMM operations on Blackwell hardware is now accessible via Python-based syntax rather than low-level CUDA C++.
• –The transition to "Triton-distributed" enables developers to write computation-communication overlapping kernels (like AllGather + GEMM) without C++ expertise.
• –PyTorch 2.11's deprecation of TorchScript cements Triton as the primary path for performance-critical production deployments.
• –Community maintenance by Meta, NVIDIA, and AMD ensures Triton remains hardware-agile as the AI accelerator landscape diversifies.