[TAO-2382][Feature] Add dinov3 evaluate action (shared core/evaluation suite)

[vpraveen-nv]

What changes are proposed in this pull request?

Mirrors the nvdinov2 evaluate action for the DINOv3 SSL family (Epic G), plus two small core/evaluation robustness fixes found during validation:

• TAO-2382 — dinov3 evaluate: registers dinov3 in ADAPTER_REGISTRY (reuses DinoV2Adapter; DinoV3PlModel(DinoV2PlModel) exposes the same teacher.backbone token dict); adds ssl/dinov3/scripts/evaluate.py (checkpoint via restore_pretrained_weights, accepts stripped teacher/student .pth or timm-format .pth/.safetensors); DINOv3EvaluateExpConfig(EvaluateConfig, EvalSuiteConfig) + evaluate_spec.yaml (KNN-only, ImageNet normalization); README table regen.
• TAO-2386 — fixes cache_tag fallback when the config field is present but None (crashed any run with evaluate.cache_dir set).
• TAO-2383 — tolerates truncated JPEGs in the eval classification dataset (ImageFile.LOAD_TRUNCATED_IMAGES = True, same policy as the mae/classification_pyt/radio dataloaders).

Why are the changes needed?

Completes embedding-quality evaluation coverage for the DINOv3 family through the shared core/evaluation suite, and fixes two crashes hit during full-scale validation (a cache_tag crash with cache_dir set, and a truncated ImageNet JPEG killing full-split extraction).

Related issues

JIRA: TAO-2381 / TAO-2382 / TAO-2383 / TAO-2386. N/A for GitHub issues.

Does this PR introduce any user-facing change?

Yes — new dinov3 evaluate action. Also: runs with evaluate.cache_dir set no longer crash, and truncated JPEGs no longer abort evaluation dataset extraction.

How was this patch tested?

Validation on a4u8g-0146 (single A100):

Check	Result
Unit tests (tests/core/evaluation + dinov3 config)	39 passed
timm feature-parity (published ViT-B/16, 256 + 768 px)	pass, 162/162 tensors remapped
dinov3 evaluate ImageNet KNN, full 1.28M index @224	knn_top1 = 82.99% (published DINOv3 ViT-B/16 k-NN ballpark)
C-RADIOv4-H parity through the same core KNN path @512	86.42% vs paper 86.59 (Δ −0.17)

pylint 10.00/10 on touched files.

Was this patch authored or co-authored using generative AI tooling?

Yes — portions were co-authored with an AI coding assistant.

Release note

Added a dinov3 evaluate action for embedding-quality (KNN) evaluation of DINOv3 checkpoints.

Checklist

• My commits are signed off per the DCO (git commit -s) — see CONTRIBUTING.md
• I have read the contributing guidelines
• The code follows the project's style, and lint and format checks pass locally
• I added or updated tests covering this change
• All tests pass locally
• I added or updated documentation (README, docstrings, docs pages, examples)
• I updated the version, changelog, or migration notes if this change requires it
• If this touches components that are optional to install, the imports are guarded
  so the package still works without them (reviewers: please verify this)
• No secrets, credentials, proprietary data, or customer data are included in this PR
• I understand this contribution is licensed under the repository's license, and that
  my commit author name and email become permanently public once merged

Notes for reviewers

• Stacked on [TAO-2175][Feature] core/evaluation unit tests (26 tests) #80 (feature/core-eval-tests); retarget to main after the stack merges. Diff vs stack top: 9 files, +239/−3.
• KNN validation ran with use_faiss=False — faiss-cpu 1.13.2 crashes in IndexFlatIP.search at scale (TAO-2387).
• ADE20K seg validation is TAO-2384 (blocked: dataset not on the devbox).
• One rebase conflict was resolved when moving to GitHub main (config/dinov3/default_config.py): kept main's newer DINOv3ExportExpConfig and added the evaluate field — worth a second look.

---

Migrated from GitLab MR nvidia-tao-toolkit/tao-pytorch!645, rebased onto GitHub main.

[tao-pr-bot]

For security reasons, CI does not run automatically on NVIDIA's runners — it must be triggered per commit.

• Comment /build to run CI on your latest commit, and make sure blossom-ci is green before merging.
• Re-run /build after each new push — CI is pinned to the head commit, so a stale run won't count.

Tip

Should this fix also ship in a release? Add a release/X.Y.Z label (matching the target release branch) and tao-cherry-pick-bot will backport it after merge.

Important

Before merging: every file needs a license header and every commit must be DCO signed-off — see the 📖 Contribution guide for details.