NVDINOv2

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

NVDINOv2 for self-supervised visual representation learning. Trains vision transformers via self-distillation (teacher-student) without labels. Produces general-purpose visual features.

Set train.pretrainedmodelpath for pretrained ViT weights.

For TAO Deploy TensorRT actions (gentrtengine), read references/tao-deploy-nvdinov2.md first. Deploy spec templates live in this skill's references/ folder with the spectemplatedeploy_*.yaml prefix.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spectemplate<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skillinfo.yaml via automlenabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spectemplate<action>.yaml to exist and parse. Use the packaged selected-action schema for automldefaultparameters, automldisabledparameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skillinfo.yaml and resolve the run override from either an explicit automlpolicy value or the user's workflow request. Use automlpolicy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automlpolicy: off for this run only. When automlpolicy: on, automlenabled: true, and both schemas/train.schema.json and references/spectemplatetrain.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skilldir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automlpolicy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Non-train actions such as inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Training Requirements

• Dataset type: image_classification
• Formats: ssl
• Monitoring metric: train_loss

Per-Action Dataset Requirements

The managed runner may source these datasets as archives because references/skillinfo.yaml declares runtime: extractedfolder. It must unpack the archive and place the resulting directory in imagesdir. For direct Docker or TAO CLI runs, extract archives before launch; never set an imagesdir field to .tar, .tar.gz, .tgz, or .zip.

Typical Spec Overrides

Data source overrides are mandatory for train and inference — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in spec_overrides.

TRAIN_IMAGES_DIR = "/workspace/data/extracted/train/images_train"
TEST_IMAGES_DIR = "/workspace/data/extracted/test/images_test"

train (mandatory data sources):

{
    "train.num_gpus": 1,
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "dataset.train_dataset.images_dir": TRAIN_IMAGES_DIR,
}

local AutoML validation / smoke run: Use this shape when the goal is to confirm Bayesian launch, metric selection, best-model choice, and checkpoint persistence on local Docker. It keeps the run representative while avoiding the much slower ViT-Large default.

{
    "wandb.enable": False,
    "model.backbone.teacher_type": "vit_s",
    "model.backbone.student_type": "vit_s",
    "model.backbone.img_size": 224,
    "dataset.batch_size": 8,
    "dataset.workers": 2,
    "train.num_epochs": 1,
    "train.checkpoint_interval": 1,
    "train.num_prototypes": 1024,
    "train.precision": "32-true",
    "train.use_custom_attention": False,
    "train.num_gpus": 1,
    "dataset.train_dataset.images_dir": TRAIN_IMAGES_DIR,
}

export (mandatory checkpoint handoff):

{
    "export.checkpoint": "<selected train/AutoML student_epoch_* checkpoint>",
    "export.onnx_file": "/path/to/results/nvdinov2.onnx",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML student_epoch_* checkpoint>",
    "model.backbone.teacher_type": "<same value used for train>",
    "model.backbone.student_type": "<same value used for train>",
    "model.backbone.img_size": "<same value used for train>",
    "train.use_custom_attention": "<same value used for train>",
    "dataset.test_dataset.images_dir": TEST_IMAGES_DIR,
}

Eval Dataset

Optional. SSL training does not use labels. Evaluation is downstream task-specific.

Important Parameters

• model.backbone.teachertype: Teacher ViT variant. Default vitl (ViT-Large).
• model.backbone.studenttype: Student ViT variant. Default vitl. Typically matches teacher.
• model.backbone.img_size: Input image size. Default 518. Higher resolution produces better features but costs more memory.
• model.backbone.patch_size: ViT patch size. Default 14.
• dataset.batch_size: Per-GPU batch size. Default 4. SSL training is memory-intensive due to dual (teacher+student) forward passes.
• train.layerwise_decay: Layer-wise learning rate decay. Important for ViT fine-tuning.
• train.clipgradnorm: Gradient clipping. Important for stable SSL training.

Multi-GPU / Multi-Node

Launch method: Lightning-managed (single python process, Lightning spawns workers).

• Strategy: auto (Lightning picks best strategy automatically)
• sync_batchnorm is always enabled — critical for SSL training with teacher-student framework
• Multi-GPU strongly recommended (4-8 GPUs) for meaningful SSL training

Multi-node env vars (set by orchestrator): WORLDSIZE, NODERANK, MASTERADDR, MASTERPORT, NUMGPUPER_NODE.

Hardware

Minimum 4 GPU(s), recommended 8 GPU(s). 40GB+ (A100 recommended) VRAM per GPU. SSL with ViT-Large teacher+student is very memory-intensive. Requires A100 40GB+ GPUs. Multi-GPU strongly recommended.

Error Patterns

No images found / archive path rejected: dataset.*.imagesdir must be an extracted directory containing images, not an archive. Extract .tar, .tar.gz, .tgz, or .zip inputs first. Managed archive-backed sources must use the directory produced by runtime: extractedfolder.

CUDA out of memory: ViT-Large teacher+student with imgsize=518 requires 40GB+ GPU memory. Reduce batchsize, img_size, or use smaller ViT variant.

Inference checkpoint has unexpected Lightning keys: For downstream inference, pass the selected AutoML run's studentepoch*.pth checkpoint, not nvdinov2modellatest.pth. The latest file is a training checkpoint and the inference loader reports unexpected keys such as state_dict, optimizer state, and scheduler state.

Export checkpoint has unexpected Lightning keys: Export also consumes the selected studentepoch.pth checkpoint. Use the full modelepoch.pth checkpoint only for resume/retrain via train.resumetrainingcheckpoint_path.

TensorRT engine passed to PyT inference: The packaged PyT nvdinov2 inference implementation only loads .pth or .tlt model paths. TAO Deploy gentrtengine builds a TensorRT engine for downstream consumers, but the PyT inference action does not run on that engine.

Separate distill action not available: The current TAO PyT CLI exposes export, inference, train, and default_specs for NvDINOv2. Do not launch or advertise a standalone nvdinov2 distill action.

AutoML metric not found: TAO's status KPI reports the final training scalar as trainloss. Use trainloss with minimize direction for AutoML selection. Some Lightning progress lines also render the same scalar as trainlossepoch; treat that as a fallback alias only, not the primary monitoring metric.

Slow convergence: SSL needs many epochs. Default 10 is for quick testing; production runs typically use 100+ epochs.

Spec Param / Parent Model Inference

Model-specific inference mappings belong in this MD file, not in config.json. Generated runners should read this section and apply the mappings with SDK helpers before createjob(). This mirrors the old microservices inferparams.py flow.

Model-specific handoff mappings:

For parentmodel or parentmodelfolder, pass the upstream train/export/AutoML child job id as parentjob_id. The SDK lists the parent result folder, filters checkpoint artifacts, and returns the selected model file or folder. Do not add these mappings back to config.json and do not patch generated runner scripts to guess checkpoint paths.

Deployment

• [tao-deploy-nvdinov2](references/tao-deploy-nvdinov2.md)