PyTorch 2.11 ships CuTeDSL for FP8 speedups

PyTorch is strategically replacing complex C++ integrations with Python-native paths that don't sacrifice hardware performance. As the fourth autotuning backend for TorchInductor, CuTeDSL specifically targets NVIDIA's H100 and B200 architectures by powering FlashAttention-4. The shift to a Python-based DSL simplifies the codebase for performance engineers while maintaining peak optimization, though some users report complexity-based slowdowns in certain graph configurations. The release also includes fallback mechanisms for older Ampere-based GPUs lacking native FP8 support.