Tao Run Dinov3 Ssl Deft

DINOv3 SSL DEFT Architecture

The application has one implementation and three interfaces:

1. SKILL.md selects a known recipe and invokes the CLI.

2. python -m nvidia_tao_ds.mining.dinov3.workflow exposes lifecycle commands.

3. nvidia_tao_ds.mining.dinov3.workflow.RefinementWorkflow is the Python API.

The controller composes leaf actions through immutable files. It never imports TAO PyTorch training in-process. It is packaged in Data Services, which owns selection, retrieval, cumulative manifests, orchestration and recovery. TAO PyTorch owns training and GRIT scoring; those modules run as subprocesses. The minimal deployment uses their existing installation inside one DS container.

The DINO base spec selects any TAO-supported vit_s, vit_s_plus, vit_b, vit_l, vit_h_plus, or vit_7b backbone. GRIT probes normalized depth quartiles and records the resolved variant and block numbers. ViT-B is the empirically studied metric; the larger-variant extension is an execution contract, not a new KPI claim.

Execution is SDK-free. One allocated DS container uses LocalRunner; distributed jobs provide an executable implementing submit, status, logs, and cancel against stage-request.schema.json. Submission must be idempotent on client_job_id, and UNKNOWN is never automatically resubmitted.

Run Trace

• input.yaml: user-facing configuration.

• resolved_run.json: immutable resolved configuration.

• release.lock.json: workflow and component versions.

• data.lock.json: checkpoint, target, benchmark, immutable source-dataset, encoder, and verified source-shard fingerprints.

• state.json: atomic current snapshot.

• events.jsonl: append-only stage history.

• jobs/: stage requests, normalized job records, and logs.

• rounds/round_NNN/: immutable stage artifacts and success markers.

• final_model.json: terminal checkpoint and training-manifest handoff.

• report.html: deterministic summary regenerated by code.

Mining treats relevance and duplication as separate fixed-encoder decisions. It may expand relevance to a configured hard floor to fill a query budget, but it does not relax the duplicate threshold. Selected source identities are unique across all weak queries in a round. data.previous_training_manifest excludes samples already used by the parent checkpoint, and cumulative manifests exclude newly selected samples in later rounds. Registered stores require a target embedding contract with the same encoder identity and dimension as the source store. Every embedding-store shard declares its byte size and SHA-256. The controller either rehashes the shards or validates a previously committed content seal, according to data.source_store_validation, then binds the verified inventory digest into data.lock.json. Image and archive locators are bound to the immutable dataset IDs, versions, and roots in data.source_payload_contract; the controller does not hash every image payload. The existing DS CLIP/SigLIP producer prepares source and target with the same immutable model/processor; no C-RADIO producer or extra installation is required. Registration with the payload contract emits the mount-bound content seal. write-target-contract validates target vectors and publishes encoder/dimension and source-inventory compatibility. Its recorded target digest is lineage only, not enforced against the workflow's current target file; regenerate on target changes. See the preparation sequence in adapter-contracts.md.

Changing a resolved configuration requires a new run directory. The controller rejects config drift on resume.

When workflow.early_stopping is present, the evaluator remains the owner of task semantics and emits the ordinary metric records. The controller combines only the explicitly named records as a weighted, direction-normalized mean. It tracks both meaningful improvement for patience and the absolute best observed score for delivery. Candidate rounds are not accepted or rejected in place; after patience is exhausted the loop stops and final_model.json points to the best evaluated checkpoint rather than the latest one. Without an evaluator, only mining and iteration-limit stop conditions are available.

Large registered pools use the audited_ann search backend. Candidate retrieval loads a persistent, sharded cuVS index and commits its row IDs as the search_candidates stage. A separate runtime then reads only those rows from the immutable float32 mining-embedding dense store, computes exact cosine scores, and applies relevance, exclusion, and duplicate rules in the normal search stage. The stage boundary isolates CUDA environments and lets reranking resume without repeating ANN retrieval. Probe count and candidate depth must exactly match a committed recall audit for that index. ANN underfill is search_budget_exhausted; only a complete exact scan can prove radius_exhausted.

For multi-task runs, normalized weakness is ranked independently within each task. A fixed targets_per_round budget is split by positive task_weights, then consumed round-robin with global sample-ID deduplication. Weights change task preference without changing total retrieval or training cost. The balanced policy preserves the configured quota of an inactive task rather than redistributing it. Materialization records the originating query task and publishes a derived training view that oversamples each provenance group to the largest group. The next round still consumes the unique cumulative manifest, so replay rows never enter lineage or source exclusion state.

Materialization writes a cumulative Parquet of unique source locators and, for balanced multi-task runs, one derived replay Parquet. A row identifies a file or an archive plus member; the controller never copies image payloads or creates symlinks. Source images and fixed mining embeddings are reused read-only. Per-run storage consists of stage Parquets and JSON, checkpoints, metrics, logs, state, and the report under run_dir.

Every loop action uses the same execution boundary. In runner mode the configured runner owns its resources; in adapter_managed mode a one-node wrapper durably owns and adopts a separate child allocation. The controller records both without knowing whether the implementation is a process, container, batch job, Kubernetes workload, or another platform primitive.

Data Services composes the complete immutable DINOv3 experiment spec and owns the round contract and final EMA-teacher publication. TAO PyTorch owns only the native dinov3 train implementation. Lightning creates DDP/FSDP workers from the spec; no workflow-specific training adapter or outer torchrun layer is used. A multi-node runner supplies the declared static gang, immutable shared inputs, and one shared output namespace. A retry cancels and restarts the full gang. Detailed launch behavior belongs to the owning DINOv3 package README.