[Upstream Breaking Change] PyTorch 2.9.1 → 2.10.0

[github-actions]

Summary

PyTorch 2.10.0 has been released with multiple backwards incompatible changes affecting DataLoader, ONNX, and torch.nn APIs. Our current pin is 2.9.1 in docker/constraints.txt.

Backwards Incompatible Changes

Critical Changes

1. ONNX Exporter: fallback=False is now the default in torch.onnx.export (#162726)

   • Previously used fallback=True by default
   • Now uses dynamo=True without fallback
   • Impact: If vLLM or build scripts use ONNX export, this will break

2. torch.nn: Removed Nested Jagged Tensor support from nn.attention.flex_attention (#161734)

   • Impact: Code using flex_attention with nested jagged tensors will fail

3. DataLoader: Removed unused data_source argument from Sampler (#163134)

   • Impact: Custom samplers that use this argument will break

4. Package Rename: pytorch-triton package renamed to triton (#169888)

   • Impact: Installation scripts and requirements.txt may break

Major Deprecations

• torch.jit: Not guaranteed to work in Python 3.14; deprecation warnings added (#167669)
• ONNX dynamic_axes: Deprecated in favor of dynamic_shapes (#165769)

New Features

• Python 3.14 support for torch.compile()
• New varlen_attn() op for ragged/packed sequences
• Intel GPU support (Panther Lake)
• LocalTensor for debugging distributed code

Files Affected

Per docs/upstream-versions.md:

• docker/constraints.txt line 1 (torch==2.9.1)

Recommended Actions

1. Audit vLLM usage for ONNX export calls
2. Check for custom samplers in vLLM codebase using data_source
3. Search for flex_attention usage with nested jagged tensors
4. Update constraints.txt:

   torch==2.10.0
   

5. Update any pytorch-triton references to triton
6. Test container build and E2E tests

Impact

HIGH - Multiple backwards incompatible changes affecting core PyTorch APIs. ONNX export behavior change may silently break inference pipelines.

Upstream References

• Release: https://github.com/pytorch/pytorch/releases/tag/v2.10.0
• Release blog: (pytorch.org/redacted)
• Full Changelog: pytorch/pytorch@v2.9.1...v2.10.0

Generated by Upstream Dependency Monitor