Fix MLU device utilities and test backend support

[SeptPonts]

What does this PR do?

Fixes #14779.

Maintainer acknowledgment:
#14779 (comment)

Complete the existing MLU backend support by:

• Registering MLU operations in the production and testing backend utility tables.
• Releasing unused MLU allocator cache in training_utils.free_memory().
• Selecting an available MLU automatically in the test utilities.

On MLU590-M9DK, the six production helpers that previously raised TypeError now succeed. After deleting a 64 MiB tensor, free_memory() releases the reserved allocator memory. The existing H3 training test passes on MLU with automatic device selection.

Before/after measurements and environment details are recorded in #14779.

Validation

Passed locally with the repository-pinned Ruff 0.9.10:

• make style and make fix-copies
• make quality
• python utils/check_copies.py
• python utils/check_dummies.py
• python utils/check_support_list.py
• python utils/check_forward_call_docstrings.py
• make deps_table_check_updated
• git diff --check

make style also proposed an unrelated, pre-existing LTX2 import-table reorder. It was excluded from this PR; make quality passed on the final three-file change.

Before/after hardware validation used MLU590-M9DK, Torch 2.12.1 and torch_mlu 1.34.1:

• Six MLU backend utility calls that raised TypeError on the parent now succeed.

• free_memory() releases the allocator cache after a deleted 64 MiB allocation: reserved memory changes from 67,108,864 bytes on the parent to zero with the patch.

• The existing H3 training test passes with automatic MLU selection:

  python -B -m pytest --assert=plain -p no:cacheprovider -q \
    tests/models/transformers/test_models_transformer_minimax_h3.py::TestMiniMaxH3TransformerTraining::test_training

The full test suite and CUDA/NPU runtime suites were not run. Hardware validation predates the final import-name sorting correction; no runtime logic changed afterward.

Self-review

Applied .ai/skills/self-review/SKILL.md to the entire three-file diff. No blocking correctness issues or unused additions were found. The known import-order lint failure has been corrected and quality checks pass.

One initialization assumption remains for maintainer review: MLU table entries expect torch_mlu registration before the utility modules are imported. Automatic extension loading supplies this in the tested environment; late extension loading with autoload disabled is not covered.

This fixes device utilities and test-backend selection. No public signatures, dtype policies or determinism settings change.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sayakpaul Following up on the discussion in #14779.

[sayakpaul]

Thanks!

[sayakpaul]

Why is this the case? is empty_cache not available generally in Torch MLU?

[SeptPonts]

Yes, empty_cache() is available in torch_mlu, and I verified it on hardware.

The guard is for the optional backend namespace: unlike the built-in torch.cuda and torch.xpu modules, torch.mlu is registered by the external torch_mlu package through PyTorch’s PrivateUse1 integration. Without the extension loaded, accessing torch.mlu.empty_cache while constructing this module-level dictionary raises AttributeError: module 'torch' has no attribute 'mlu'. This prevents the shared utilities from being imported even for CPU/CUDA tests, before any MLU operation is requested.

I followed the guarded-access pattern already used for Neuron’s device_count and synchronize entries in src/diffusers/utils/torch_utils.py. The same reasoning applies to the device_count comment below.

[sayakpaul]

Same as above.