The Wire · Showcase
PYTORCH 2.13 SHIPS WITH FLEXATTENTION ON APPLE SILICON, NCCL2 BACKEND LANDS IN-TREE
By RepoJournal · Filed · About PyTorch
PyTorch's biggest release in months brings native Apple Silicon acceleration, refactors distributed training infrastructure, and ships quantized embedding support across CPU and accelerators.
PyTorch 2.13.0 is live [1], and the marquee feature is FlexAttention landing on Apple Silicon (MPS) with up to 12x speedup over SDPA on sparse patterns, making M-series deployments finally competitive for attention-heavy workloads. Simultaneously, the NCCL2 backend port [2] begins shipping in-tree as a c10d-native backend selected via init_process_group(backend='nccl2'), replacing the external torchcomms dependency and centralizing distributed training under one roof. ExecutorCh rounds out the release with bf16 quantized embedding support on CPU [3], bf16/fp16 activations in SDPA [4], and XNNPACK delegation fixes for even-kernel same-padding convolutions [5], unblocking deployment patterns that previously fell back to slow paths. The profiler lands a critical CUPTI timestamp callback fix [6] that eliminates per-record clock conversions, and the test suite gains hardware-independent SAC-ILP tests [7] to stop flakiness across GPU models. Test-infra ships release 2.13 go-live [8], advancing stable and candidate to 2.13.0 together, while also introducing a new 'infra_issue' AI verdict [9] to separate real CI failures from code problems.
Action items
- → Test FlexAttention on Apple Silicon targets to unlock 12x sparse attention speedups pytorch/pytorch [plan]
- → Begin planning NCCL2 backend migration from external torchcomms pytorch/pytorch [plan]
- → Ship bf16 embedding quantization to production if you target CPU inference pytorch/executorch [plan]
- → Monitor 2.13 stability in first 48 hours before broad rollout pytorch/pytorch [monitor]
References
- [1] PyTorch 2.13.0 Release ↗ pytorch/pytorch
- [2] [c10d][nccl2] Port torchcomms NCCL backend foundation (utils, CUDA API, batch) (#188582) pytorch/pytorch
- [3] Support bf16-out quantized embeddings on CPU ↗ pytorch/executorch
- [4] Support bf16/fp16 activations in CPU SDPA (#20611) ↗ pytorch/executorch
- [5] Delegate even-kernel 'same'-padding convs via a quantized static pad (#20553) ↗ pytorch/executorch
- [6] [profiler][cupti] Engage the approx-clock timestamp callback via the per-subscriber attribute (#189168) pytorch/pytorch
- [7] Make test_sac_ilp hardware-independent by pinning device datasheet (#189278) pytorch/pytorch
- [8] Release 2.13 go live. Update release matrix (#8261) pytorch/test-infra
- [9] [autorevert] Add infra_issue AI advisor verdict (treated like not_related) ↗ pytorch/test-infra
FAQ
- What changed in PyTorch on July 9, 2026?
- PyTorch's biggest release in months brings native Apple Silicon acceleration, refactors distributed training infrastructure, and ships quantized embedding support across CPU and accelerators.
- What should PyTorch teams do about it?
- Test FlexAttention on Apple Silicon targets to unlock 12x sparse attention speedups • Begin planning NCCL2 backend migration from external torchcomms • Ship bf16 embedding quantization to production if you target CPU inference
- Which PyTorch repositories shipped on July 9, 2026?
- pytorch/pytorch, pytorch/executorch, pytorch/test-infra