chore(deps): bump torch from 2.13.0 to 2.14.0

[dependabot]

Bumps torch from 2.13.0 to 2.14.0.

Release notes

Sourced from torch's releases.

PyTorch 2.14.0 Release Notes

• Highlights
• Backwards Incompatible Changes
• Deprecations
• New Features
• Improvements
• Bug fixes
• Performance
• Documentation
• Security
• Developers

Highlights

For more details about these highlighted features, you can look at the release blogpost. Below are the full release notes for this release.

Backwards Incompatible Changes

torch.nn

• torch.nn.LinearCrossEntropyOptions no longer accepts acc_policy="balanced"; use "compact" instead (#188283)

  The "balanced" policy was removed because "compact" provides the same weight-gradient accumulation precision with lower memory use on CUDA, already uses the equivalent scratch layout for mixed-precision inputs on other devices, and was never selected by "auto". Constructing the options with acc_policy="balanced" now raises ValueError: invalid acc_policy: 'balanced'; expected one of 'auto', 'accurate', 'compact'.

  Before:

  options = torch.nn.LinearCrossEntropyOptions(acc_policy="balanced")
  loss = torch.nn.functional.linear_cross_entropy(
      input, linear_weight, target, options=options
  )

  After:

  options = torch.nn.LinearCrossEntropyOptions(acc_policy="compact")
  loss = torch.nn.functional.linear_cross_entropy(

... (truncated)

Commits

• 2b3ec34 [release/2.14] Import SDPAParams in test_transformers to fix lint (#194970)
• 08187d9 [cuDNN] Add guards for cuDNN SDPA decode (#194963)
• 8ceea97 Pin cython < 3.3.0 for the Windows Triton wheel build (#194931)
• 99ecebc [Cherry-pick][release/2.14] [inductor] Fix loop-local load CSE lifetime (#194...
• ec283a7 Bump the Python 3.15 numpy pin to 2.5.2 (#194821)
• 65890f3 Fix docker-release validate job to use the channel matching the pushed image ...
• 1682388 Fix macOS py3.15 wheel builds: pin Cython < 3.3.0 and bump the cp315 numpy pi...
• 9724418 Fix Windows py3.15 builds: constrain Cython < 3.3.0 and bump the cp315 numpy ...
• f1b7554 [MPS] Fix pin_memory() recycling buffers still in use by the GPU (#194662)
• 9f205f7 [MPS] fail loudly on large reductions (#194661)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[imran-siddique]

Reviewed ad94509.

The existing CI and Smoke runs pass, but neither installs requirements-lora.txt. The WCM job installs requirements.txt and runs offline demos, so those green checks do not validate torch 2.14.0 with the pinned Transformers, PEFT and Accelerate versions. Holding approval until the complete LoRA requirements resolve and an offline adapter creation/save/load smoke test runs against this pin. This is a coverage gap, not an observed PyTorch regression.

[imran-siddique]

Reviewed ad94509.

Closed the coverage gap from my previous review with an isolated installation of the exact LoRA requirements: torch 2.14.0, huggingface_hub 1.29.0, Transformers 5.16.1, PEFT 0.20.0, Accelerate 1.14.0, safetensors 0.8.0 and model-signing 1.1.1. Using a locally generated tiny Llama model/tokenizer with Hub access disabled, exercised the PR's real_lora_custody.train_lora for two gradient steps, confirmed nonzero learned LoRA B weights, encrypted/decrypted the saved adapter with inventory verification, rejected a wrong key, then reloaded through run_inference successfully. This is an offline compatibility test, not a hosted model download or GPU benchmark. Existing hosted CI also passes apart from the review gate.