●Change8

Migrating to PyTorch v2.7.1

Version v2.7.1 introduces 2 breaking changes. This guide details how to update your code.

Released: 6/4/2025

2
Breaking Changes
8
Migration Steps
9
Affected Symbols

⚠️ Check Your Code

If you use any of these symbols, you need to read this guide:

torch.compiletorch.autocastflex_attentiontorch.distributed.barriertorch.mkldnn_max_pool2dtorch.nn.functional.grid_sampletorch.profilertensor.viewCompositeImplicitAutograd

Breaking Changes

●Issue #1

If you were relying on specific error logging behavior within `torch.compile`, the logging implementation has been improved, which might change the format or verbosity of logged errors.

✓Solution

Review any custom error handling or logging configurations around `torch.compile` calls and adjust them if necessary based on the new logging output.

●Issue #2

If you were using mutable custom operators with `torch.compile`, they are now marked as cacheable, which could potentially alter caching behavior if the mutability was previously causing issues or unexpected re-compilations.

✓Solution

Verify that marking mutable custom operators as cacheable does not introduce correctness issues in your compiled graphs. If issues arise, you might need to refactor the custom operator to be immutable or use alternative compilation strategies.

Migration Steps

  1. 1
    Update your PyTorch installation to the new version.
  2. 2
    If you are using HuggingFace LLM models with `torch.compile`, test your workloads thoroughly to benefit from reduced cudagraph re-recording and improved compatibility.
  3. 3
    If you encounter crashes when using `torch.autocast` alongside `torch.compile`, ensure your environment is updated, as a fix for exceptions raised inside `autocast` has been implemented.
  4. 4
    If you use `einops`, ensure you are using a version compatible with this release, or be aware that a workaround for older versions has been implemented.
  5. 5
    If you are using distributed training, monitor for hangs or unexpected behavior, especially when using non-blocking APIs with NCCL 2.26, as workarounds have been added.
  6. 6
    If you are developing on macOS with Clang 17, ensure your build environment is correctly configured, as compilation errors have been resolved.
  7. 7
    If you observe incorrect results on MPS devices when using binary operations involving wrapped scalars, test your MPS workloads.
  8. 8
    If you are using ONNX export/import, verify that CompositeImplicitAutograd ops are correctly preserved during decomposition.

Release Summary

This maintenance release focuses on fixing regressions and silent correctness issues across torch.compile, Distributed, and Flex Attention, while also improving wheel sizes and platform-specific compatibility for MacOS, Windows, and XPU.

Need More Details?

View the full release notes and all changes for PyTorch v2.7.1.

View Full Changelog