[inureyes]

Hugging Face DINOv2 checkpoints previously had no shared encoder or standalone runtime in mlxcel, which blocked the dense-model families that depend on the backbone. This adds a reusable standard/register-token encoder and exposes image feature extraction through mlxcel predict.

The encoder loads public Hugging Face standard and register checkpoints, remaps and validates their weights, supports boolean patch masks, preserves register-token ordering, returns normalized intermediate layers, and implements rectangular positional interpolation in both upstream modes. Shared consumers can construct a backbone from a scoped parent checkpoint without consuming the parent weight map. The giant preset uses its SwiGLU feed-forward layout.

The standalone predictor reads preprocessor_config.json and reproduces the slow Hugging Face BitImageProcessor path: RGB conversion, shortest-edge resize, Pillow-compatible bicubic uint8 rounding, center crop, rescaling, and normalization. It writes last_hidden_state, pooler_output, and spatial-only hidden_patch_tokens. Architecture and operator documentation describe both supported checkpoint types.

Pinned full-tensor CPU f32 fixtures cover facebook/dinov2-small and facebook/dinov2-with-registers-small after real checkpoint preprocessing. Tests also cover predictor preprocessing, positional interpolation literals, [4, 11] intermediate outputs, register exclusion, f16 masking, scoped loading, allocation limits, invalid public-layer inputs, and sanitizer behavior. Checkpoint-driven image, token, architecture, and intermediate-output sizes are bounded before expensive allocations; attention uses the shared safe dispatch.

Validation:

• cargo test -p mlxcel --profile test-fast --features rocm --lib dinov2 -- --nocapture --test-threads=1: 13 passed, 0 failed, 0 ignored. The explicit CPU f32 reference tests produced relative RMS values from 1.197926287e-7 for preprocessing through 4.275870861e-6 for model outputs.
• cargo clippy -p mlxcel --profile test-fast --features rocm --lib --tests -- -D warnings
• cargo build -p mlxcel --profile test-fast --features rocm --bin mlxcel
• Guarded MLXCEL_DEVICE=gpu CLI predictions on gfx1151 for both checkpoints wrote exactly meta.json and result.safetensors. Standard-checkpoint tensor RMS values were 2.844445329e-6, 2.473357604e-6, and 2.845811506e-6; register-checkpoint values were 4.548628530e-6, 2.131596900e-6, and 4.596212478e-6 for last hidden state, pooler output, and patch tokens respectively.
• mlxcel arch --json lists dinov2 with both model types and the predict/image/embeddings capability.
• cargo fmt --all -- --check
• python3 scripts/ci/check_env_var_docs.py
• python3 scripts/ci/check_binary_assets.py
• git diff --check
• Independent implementation and security/performance reviews completed with all findings resolved.

Closes #2350