PyTorch 2.14 lands Mac linear algebra and a Python 3.15 trap

PyTorch 2.14 shipped on September 15 with 2,995 commits from 487 contributors since 2.13. Three things in it matter if you run models rather than research them: Apple Silicon finally got native linear algebra, fault tolerance stopped being an NCCL implementation detail, and the new Python 3.15 wheels will happily import and then refuse to compile.

What actually happened in PyTorch 2.14

The Apple Silicon work is the headline for anyone whose dev machine is a Mac. MPS previously leaned on Apple's MPSGraph or fell back to CPU for anything past the basics, which made mixed CPU/MPS round-trips a standard source of slowdown. This release adds native Jacobi-kernel SVD, eigh and lstsq for float32 and complex64, a faster Cholesky panel factorization (roughly 1.2–2.8x), and hand-written Metal kernels for lu_factor and lu_solve — where the submitter measured over 100x on small batched matrices and 2–9x on larger single ones. A second prefill attention kernel, built on Apple's Metal Performance Primitives in macOS 26.2, benchmarks 2–4x faster than the previous one. There is also a routing fix for single-token decode through F.linear, a shape that had been falling off the fast path and costing an 8.5x slowdown on bf16 and fp16.

On the distributed side, the rewritten NCCL backend from torchcomms lands in-tree as a preview nccl2 c10d backend and becomes the default in 2.15. Fault tolerance is now a first-class c10d concept: a process group can be reconfigured in place after a rank failure instead of being torn down and restarted, and the Flight Recorder works across backends rather than only NCCL.

Now the trap. PyTorch 2.14 publishes Python 3.15 wheels, including the free-threaded 3.15t build — but Python 3.15 support is eager-only. Calling torch.compile under 3.15 raises a RuntimeError rather than silently falling back. Those wheels are also not on PyPI; you pull them from download.pytorch.org.

Why this matters for your business

The Apple Silicon numbers change where work can happen. Fine-tuning a small model, running an eval sweep, or batch-processing embeddings on the laptop you already own is now a real option instead of a demo — and every hour of that is an hour not billed by a GPU host. For a small team, the dev-loop cost is the cost that compounds.

The Python 3.15 note is the one to act on this week. The failure is loud by design, which is good, but it will still surprise whoever bumps the base image and then watches a training job die with a RuntimeError instead of a slow run. Pin your runtime deliberately: if anything in your pipeline touches torch.compile, stay on 3.14 until 2.15. TorchScript deprecation warnings are also now visible, so expect log noise from older code.

Sources: PyTorch 2.14 release blog, PyTorch v2.14.0 release notes.