●Change8

Migrating to PyTorch v2.9.1

Version v2.9.1 introduces 1 breaking change. This guide details how to update your code.

Released: 11/12/2025

1
Breaking Changes
4
Migration Steps
10
Affected Symbols

⚠️ Check Your Code

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

torch.nn.functional.conv3dtorch.compiletorch.bmmtorch.distributed.gathertorch.ops.inductor.mark_static_addressCUDNN_ATTENTIONfused_scaled_matmul_reduce_scattertorch.cuda.set_float32_matmul_precisionbytecode_transformationget_free_symbol_uses

Breaking Changes

●Issue #1

If you were using F.conv3d with bfloat16 inputs and experienced significant memory regressions after PyTorch 2.9.0, this release provides a fix, but it requires an external dependency update.

✓Solution

To resolve the memory regression, ensure you have installed the nvidia-cudnn package version 9.15 or newer via pip (e.g., pip install nvidia-cudnn==9.15+).

Migration Steps

  1. 1
    1. Update your PyTorch installation to the latest version.
  2. 2
    2. If you are using F.conv3d with bfloat16 inputs and noticed memory issues since PyTorch 2.9.0, install the required external dependency: 'pip install nvidia-cudnn>=9.15'.
  3. 3
    3. Review your usage of torch.compile. While most changes are fixes, monitor performance and graph breaks, especially if you use Gemma models or vLLM integration, as underlying compilation logic has been updated.
  4. 4
    4. If you were previously suppressing warnings related to new TF32 APIs, these warnings have been fixed and may no longer appear, potentially changing the console output during startup.

Release Summary

This maintenance release addresses critical regressions in PyTorch 2.9.0, specifically fixing memory issues in 3D convolutions, Inductor compilation bugs for Gemma/vLLM, and various distributed and numeric stability fixes.

Need More Details?

View the full release notes and all changes for PyTorch v2.9.1.

View Full Changelog