pytorch/pytorch v2.13.0: PyTorch 2.13.0 Release

Key points

• <tr><td><strong>FlexAttention</strong> lands on Apple Silicon (MPS), with up to ~12x speedup over SDPA on sparse patterns, and gains a deterministic backward path on CUDA for reproducible gradient computation.</td></tr>1
• <tr><td><strong>CuTeDSL "Native DSL" backend</strong> gives Inductor a second high-performance code path (alongside Triton) for key GPU operations, with faster compilation.1
• <tr><td><strong><code>nn.LinearCrossEntropyLoss</code></strong> combines the final prediction and loss computation to cut peak GPU memory by up to 4x for large-vocabulary language model training.</td></tr>1
• <tr><td><strong>Python 3.15 wheel support</strong> for PyTorch on Linux via the pytorch repository index, including builds compatible with free-threaded 3.15t.</td></tr>1

Sources (1)

• [1]pytorch/pytorch v2.13.0: PyTorch 2.13.0 Release
  GitHub: pytorch/pytorch · Jul 8, 05:39 PM

  <tr><td><strong>torchcomms</strong>, a new communications backend for PyTorch Distributed, improves fault tolerance, scalability, and debuggability for large-cluster training.</td></tr>

  <tr><td><strong>FlexAttention</strong> lands on Apple Silicon (MPS), with up to ~12x speedup over SDPA on sparse patterns, and gains a deterministic backward path on CUDA for reproducible gradient computation.</td></tr>

Extractive summary: sentences quoted from the sources.