PyTorch 2.13 is released with 3,328 commits from 526 contributors. Key highlights include FlexAttention landing on Apple Silicon (MPS) with up to ~12x speedup over SDPA on sparse patterns, a new CuTeDSL 'Native DSL' backend for Inductor as an alternative to Triton for GEMM and RMSNorm, and nn.LinearCrossEntropyLoss that fuses linear projection and loss computation to cut peak GPU memory by up to 4x for large-vocabulary LM training. Distributed training gains the new torchcomms backend for improved fault tolerance and debuggability, plus FSDP2 now supports overlapping reduce-scatter and all-gather via separate process groups. Python 3.15 wheel support (including free-threaded 3.15t) is available on Linux. Platform updates include ROCm AOTriton 0.12b, Armv9-A torch.compile targeting, and Intel XPU device telemetry APIs. Breaking changes include removal of named tensors and the Bazel build system.
