AutoGineer
Generated banner for Tao Port Huggingface Model
Generated logo for Tao Port Huggingface Model

Tao Port Huggingface Model

NVIDIAOfficialCreated by NVIDIA Corporation

  • Skill
  • Workflow
  • Software development
  • Apache-2.0
  • Medium risk
Download ZIP

From the author

Integrate a HuggingFace Computer Vision model into the NVIDIA TAO Toolkit ecosystem (tao-core config, tao-pytorch trainer, tao-deploy TensorRT pipeline). Use when the user asks to "integrate a HuggingFace model into TAO", "add an HF model to TAO Toolkit", "wire a HuggingFace ViT/DETR/ SegFormer into tao-pytorch", "build a TAO trainer + deploy pipeline for an HF CV model", or pastes a HuggingFace model URL/ID and wants it turned into a TAO model. Covers the full 7-phase loop: prerequisites check, HuggingFace inspection and validation, codebase exploration, tao-core configuration and native trainer implementation, ONNX export plus TensorRT deploy integration, packaging and L0 testing, container-based end-to-end validation, and (conditional) accuracy/latency tuning. Supports classification, object detection, semantic / instance / panoptic segmentation, zero-shot detection, and depth estimation.
— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a structured, seven-phase workflow for porting HuggingFace Computer Vision models into the NVIDIA TAO Toolkit. It guides the user through inspecting the model, configuring the TAO core, implementing the trainer, and setting up the deployment pipeline for TensorRT. The process emphasizes a local-only development loop, requiring the user to manage four independent repository clones and execute all validation, testing, and training tasks within locally prepared TAO Toolkit containers.

What it does

  • Validates HuggingFace model compatibility for computer vision tasks.
  • Configures TAO core and implements native trainers for PyTorch.
  • Integrates ONNX export and TensorRT deployment pipelines.
  • Automates packaging and L0 unit testing for model components.
  • Performs end-to-end validation within containerized environments.
  • Provides optional tuning for accuracy, latency, and model size.

When to use it

  • When you need to bring a custom HuggingFace computer vision model into the TAO Toolkit.
  • When you require a full training-to-deployment pipeline for a new model architecture.
  • When you need to verify model performance and exportability to TensorRT.
  • When you are working in a local environment with NVIDIA GPU support.

Advantages

  • Ensures consistency across training, export, and deployment stages.
  • Provides a rigorous, step-by-step debugging and testing loop.
  • Maintains environment isolation using containerized workflows.
  • Supports a wide range of computer vision tasks including detection and segmentation.

Usage examples

“Integrate a HuggingFace model into TAO.”
The agent initiates the Phase 0 prerequisite check to verify your environment and prepare the necessary TAO Toolkit container images.
“Wire a HuggingFace SegFormer into tao-pytorch.”
The agent begins the codebase exploration phase to map the model to the appropriate TAO reference architecture and prepare the configuration.

Install

  1. Download the skill as a ZIP:

    https://mp-staging.autogineer.com/agent-skills/nvidia/tao-port-huggingface-model/download

  2. Unzip it into the cross-client skills folder of your project:

    mkdir -p .agents/skills && unzip tao-port-huggingface-model.zip -d .agents/skills

Documentation by NVIDIA Corporation

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).

<!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->

TAO-HF Integration Skill

Integrate a HuggingFace (HF) Computer Vision model into the NVIDIA TAO Toolkit ecosystem. Work the phases iteratively — not purely linearly — via a build → test → debug → fix → retest loop at every step: when something fails, diagnose and fix before moving on; when it passes, move to the next step.

This SKILL.md is the workflow coordinator. Each phase has a dedicated references/phase-N-*.md with the full step-by-step content, code, docker invocations, and gates. Read the matching reference at the start of each phase — the summaries below are not sufficient.

Local-Only Rule

All work is strictly local. Do NOT push/commit/branch on any remote (GitLab, GitHub, HuggingFace), create merge/pull requests or issues, or upload/publish Docker images to any registry or artifact store. You may only read/clone from remotes — all edits, Docker builds, and test runs stay on the local machine.

Submodule Override Strategy

The user clones the four TAO repos (tao-core, tao-pytorch, tao-deploy, tao-dataservices) independently into one working directory. The tao-core/ submodule nested inside each repo points to the original unmodified commit; modifications only exist in the top-level tao-core/. Always install from the top-level tao-core/, never <repo>/tao-core/ — the nested submodule silently ignores all modifications. Override rules: (1) mount the working directory -v $(pwd):/workspace; (2) pip install /workspace/tao-core FIRST, before tao-pytorch/tao-deploy; (3) PYTHONPATH top-level tao-core first, e.g. -e PYTHONPATH=/workspace/tao-core:/workspace/tao-pytorch. See references/cross-cutting.md for the directory tree.

Execution platform

Every test, smoke run, and end-to-end validation executes inside a locally prepared TAO Toolkit container (tao-pytorch-base:latest, tao-deploy-base:latest, optionally tao-dataservices-base:latest — all from Phase 0). The platform skills own how to run them; this skill specifies what. Default platform: local-docker. Phase 0 delegates the driver / CUDA / NCT preflight to tao-setup-nvidia-gpu-host. See references/cross-cutting.md for the authoritative-skill table, bind-mount rationale, and canonical docker-run flag set.

Phase Map

PhaseGoalReference
0Prerequisites + TAO Toolkit images + local image tagsphase-0-prereqs.md
1Inputs, HF-inspection container, validate model + datasetphase-1-inspection.md, hf-inspection.md
2Closest existing TAO reference modelphase-2-codebase.md, task-type-guide.md
3tao-core config + tao-pytorch trainer / eval / inferencephase-3-implementation.md, tao-patterns.md, repo-structure.md
4ONNX export + tao-deploy TRT engine / inference / evalphase-4-deploy.md
5Packaging (console_scripts) + L0 testsphase-5-packaging.md
6Container testing + end-to-end validationphase-6-container-tests.md, docker-patterns.md
7(conditional) Accuracy / latency / size tuningphase-7-optimization.md

Cross-cutting refs: workflow-consistency.md (CLI flow, config field paths, cross-phase dependencies); cross-cutting.md (platform, isolation, module pitfalls, debugging).

IMPORTANT — Continuous Execution Through Phase 6: do NOT stop after Phases 3–5 to wait for the user to run tests. Phase 6 is mandatory — not complete until tests pass inside the containers and the end-to-end pipeline is validated.

Development Loop

At every step: write code → test immediately (import check, unit test, or dry-run) → if it fails, read traceback → diagnose → fix → retest; if it passes, move on. Do NOT accumulate untested code — testing only at the end compounds bugs.

Debugging Playbook

When something fails, consult the symptom → likely-cause → fix table in references/cross-cutting.md before trying random fixes — it covers ModuleNotFoundError, BACKBONE_REGISTRY KeyError, shape mismatch, NaN loss, ONNX/TRT build failures, TRT-vs-PyTorch accuracy gaps, OOM, DDP hangs, checkpoint load failures, and stale-submodule config issues.

Environment Isolation Strategy

All Python work runs inside Docker containers — no host venvs, no pip installs into host Python (the host needs only Docker, from tao-setup-nvidia-gpu-host). Three contexts: A (Phase 1 HF inspection in tao-hf-inspect, python:3.12-slim fallback), B (Phase 3/4/6 smoke/L0/e2e in the prepared container, source via pip install /workspace/tao-core && python setup.py develop), C (host-bind-mount scratch). See references/cross-cutting.md for the contexts in full and the four numbered rules verbatim (--check-only host packages; Phase 1 --user $(id -u):$(id -g) vs. root; HOME=/workspace/PIP_USER=1 fallback; distro package-manager list; root:root trade-off; tao-hf-inspect cleanup).

Phase 0 — Prerequisites Check

Goal: verify Python 3.10+ and git; delegate the driver / CUDA / Docker / NVIDIA Container Toolkit host check to tao-setup-nvidia-gpu-host; verify NGC docker login for nvcr.io. Then ask the user for the TAO Toolkit image references (tao-pytorch, tao-deploy, optionally tao-dataservices), pull, and prepare local tags tao-pytorch-base:latest, tao-deploy-base:latest, tao-dataservices-base:latest for later phases — preparation removes the pre-installed released TAO packages so the user's /workspace/... clones install/load via pip install /workspace/tao-core && python setup.py develop. Hard stop on any failed check. Required user inputs: the image references + credentials (NGC login, HF_TOKEN). Full commands, prompt wording, and per-image Dockerfile snippets: the Phase 0 reference.

Gate: all prerequisite checks pass; the user supplied the required image references; tao-pytorch-base:latest and tao-deploy-base:latest exist locally; tao-dataservices-base:latest exists if dataservices work is anticipated.

Phase 1 — Information Gathering & Validation

Goal: decide whether to proceed at all. Gather credentials, locate/clone the four TAO repos, create a consistent working branch, launch the tao-hf-inspect container (Context A), validate the HF model is CV with a supported pipeline_tag, extract config + state-dict schema, sanity-check ONNX export, clean up. Full steps: Phase 1 references.

Reject if: pipeline_tag is NLP / audio / LLM (non-CV); AutoConfig raises; or ONNX export fundamentally cannot work (no rewrite path).

Gate: all 4 TAO repos located/cloned with a consistent branch; pipeline_tag confirmed CV; model_type, image_size, hidden_size, num_labels extracted; state-dict keys documented + HF→TAO remapping plan drafted; ONNX export sanity check passed (or failure understood); user confirmed model_short_name + task type. (Full checklist: phase-1-inspection.md.) Present findings and get user confirmation first.

Phase 2 — Codebase Exploration

Goal: find the closest existing TAO reference model for the detected pipeline_tag, read its implementation across tao-core / tao-pytorch / tao-deploy, and decide whether the backbone exists in backbone_v2/ or is new.

The HF pipeline_tag → TAO reference model mapping (classification → classification_pyt, detection → dino/rtdetr, segmentation → segformer, instance → mask2former, panoptic → oneformer, zero-shot → grounding_dino, depth → mono_depth) drives everything downstream (config, architecture, loss, ONNX shape, TRT builder, deploy classes, metrics, dataset format). See the Phase 2 references for the full reference list (12 files per model), the backbone_v2/ and tao-dataservices coverage checks, and per-task architecture.

If a new backbone is needed, decide the strategy (timm wrap > re-implement > HF black-box wrap) before Phase 3 — it changes weight loading, ONNX export, deploy. Never dual-inherit from transformers.PreTrainedModel and BackboneBase (metaclass conflict — compose instead).

Gate: reference TAO model identified + all 12 reference locations read; task-type implications understood (architecture, loss, ONNX outputs, deploy classes, metrics, dataset); backbone coverage decided (reuse / wrap timm / new); dataservices coverage checked. (Full checklist: phase-2-codebase.md.)

Phase 3 — TAO Core Configuration & Native Implementation

Goal: write the tao-core config schema + the tao-pytorch trainer / native inference / evaluation, smoke-testing between steps. (<model_name> = snake_case short-name; <ModelName> = PascalCase.)

Steps 1–7 (each builds on the previous, smoke-test between): tao-core config (1), tao-pytorch trainer (2), multi-GPU/multi-node (3), native inference → result.csv (4), native evaluation → results.json (5), MLOps for training and eval/infer → status.json (6–7). The ExperimentConfig(CommonExperimentConfig) must contain model, dataset, train, evaluate, inference, export, gen_trt_engine, quantize. All ??? fields are MISSING (user supplies via YAML/CLI); the augmentation.mean/std, model.head.in_channels, checkpoint-name, and onnx_file matches are in the checklist below.

Full per-step bodies, code, the canonical experiment_spec.yaml, and smoke-test commands: the Phase 3 references.

Gates: Step 1 — ExperimentConfig imports cleanly in-container; Step 2 — build_model(cfg) runs + PLModel instantiates in-container; Phase 3 — all 7 steps complete, smoke tests pass, no missing __init__.py.

Phase 4 — Export, Deployment & TensorRT Integration

Goal: ONNX export from tao-pytorch, then TRT engine builder + inference + evaluation in tao-deploy reusing the tao-core ExperimentConfig.

Steps 8–11: ONNX exporter (8 — task-specific input/output names, batch_size=-1 ⇒ dynamic batch); TRT engine builder (9 — subclass EngineBuilder or reuse ClassificationEngineBuilder; write specs/{gen_trt_engine,inference,evaluate}.yaml, same ExperimentConfig schema, augmentation.mean/std MUST match training); TRT inference → result.csv (10); TRT eval → results.json (11). See the Phase 4 reference for full code and the Phase 3+4 gate (3 in-container checks: imports, model build + forward, ONNX round-trip).

Module pitfalls: tao-pytorch and tao-deploy have separate hydra_runner and monitor_status — use the deploy versions in deploy scripts. ExperimentConfig comes from nvidia_tao_core in both (same schema/field paths).

Phase 3+4 gate: all three in-container checks pass (tao-pytorch imports + model + ONNX export; tao-deploy imports).

Phase 5 — Packaging & L0 Testing

Goal: register the model as a console_script in both repos and add unit tests.

Steps 12–15: register '<model_name>=...:main' in console_scripts of tao-pytorch/setup.py (12) and tao-deploy/setup.py (13, creating the deploy entrypoint/<model_name>.py via entrypoint_hydra); deploy L0 tests (14); trainer L0 tests — Trainer(..., fast_dev_run=True) + @pytest.mark.cv_unit @pytest.mark.<model_name> (15). See the Phase 5 reference for exact entry-point strings, code, and L0 test file lists.

Gate: entrypoints registered; pytest files exist and follow the marker convention. Do NOT stop — go directly to Phase 6.

Cross-Phase Data Flow & Consistency Verification

Before Docker testing, verify the chain train → export → gen_trt_engine → inference / evaluate (the *_model_latest.pth → .onnx → .engine artifact flow + the config fields each stage reads/writes — full diagram in references/cross-cutting.md).

Consistency checklist (verify before proceeding): self.checkpoint_filename → the *_latest.pth name evaluate.checkpoint / export.checkpoint reference; augmentation.mean/std identical across training spec, inference.yaml, evaluate.yaml, engine-builder preprocess_mode; ONNX input_names=['input'] / output_names=['output'] (detection/instance-seg use task-specific names); export.input_width/input_height match dataset.img_size; model.head.in_channels matches model_params_mapping.py; classes.txt at dataset.root_dir readable by both repos; all __init__.py exist (incl. scripts/__init__.py for get_subtasks() via pkgutil). Full paths: workflow-consistency.md.

Phase 6 — Container Testing & End-to-End Validation

Mandatory — start immediately after Phase 5. All TAO models ship as Docker images; code that only works outside a container is incomplete. Testing runs directly inside the TAO Toolkit container — no image build in the loop: mount → install source (setup.py develop) → run pytest / pylint / pydocstyle / flake8 directly. Use vanilla commands, NOT the ci/run_functional_tests.py / ci/run_static_tests.py wrappers (internal-mirror-only; public github.com/NVIDIA-TAO/ mirrors have no ci/ dir).

Steps 16–25: verify local image tags exist (16); unit tests for tao-core / tao-pytorch (-m cv_unit, --shm-size=16G) / tao-deploy (17–19); lint (20); wheels (21); end-to-end — train dry-run + export in one tao-pytorch session, then gen_trt_engine + inference + evaluate in one tao-deploy session (same session critical — --rm discards installs) (22); cross-check native vs TRT (23); debug shells (24); optional release images (25).

Full commands (every docker run, per-container env-vars, exact pytest / lint invocations + full-suite variants, the train/export/gen_trt_engine/inference/evaluate one-liner with all CLI overrides, the ci/ note, the fix-and-retest loop) and build scripts / runner patterns: see the Phase 6 references.

Phase 6 gate (Done criteria): tao-core / tao-pytorch / tao-deploy unit tests pass in their containers; static tests pass (or only legacy lint warnings); wheels build; end-to-end <model_name>_model_latest.pth → model.onnx → model.engine → non-empty result.csv + results.json; native vs TRT agree within tolerance.

Phase 7 — Optimization & Tuning (conditional)

Enter only if Phase 6 passes but accuracy / latency / size needs improvement. Ask the user for target metrics first.

Diagnostic categories: accuracy too low; TRT-vs-native gap; training too slow; inference too slow. Techniques: Step 27 — hyperparameter tuning; Step 28 — INT8 quantization (PTQ via torchao / modelopt, TRT INT8 + calibration); Step 29 — channel pruning + retrain; Step 30 — knowledge distillation; Step 31 — resolution tuning (TAO interpolates ViT positional embeddings automatically). See the Phase 7 reference for each category's checks, config blocks, YAML overrides, decision tree, and rationale.

Argument

$ARGUMENTS

If provided, interpret $ARGUMENTS as the HuggingFace model ID or URL to start Phase 1. If credentials or model short-name are not included, ask the user for them before proceeding.

Files

references/tao-patterns.md

<!-- Copyright (c) 2026, NVIDIA CORPORATION. All rights reserved. Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0 Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. -->

TAO Toolkit Code Patterns

Use this reference only when the parent SKILL.md points here for the current task. If this file conflicts with current SKILL.md, skill_info.yaml, schemas, or platform/model skills, the current authoritative source wins.

Contents

    1. Config Dataclasses (tao-core)

    • Field types

    • Base classes for sub-configs

    • Top-level ExperimentConfig pattern

    • Model parameters mapping

    1. Backbone (tao-pytorch)

    • Required abstract methods on BackboneBase

    • Already-registered backbones in backbone_v2/

    1. build_model() Pattern (tao-pytorch)

    • StateDictAdapter for cross-model checkpoint loading

    1. PyTorch Lightning Module (tao-pytorch)

    • TAOLightningModule provides

    • Constructor pattern

    • Callbacks (configure_callbacks)

    • Status logging

    • Checkpoint save identifier

    1. Script Entrypoints (tao-pytorch scripts)

    • Decorator stack (all scripts use this)

    • Train script (initialize_train_experiment)

    • CLI Entrypoint (model-level CLI)

    1. ONNX Export (tao-pytorch)

    1. TensorRT Engine Builder (tao-deploy)

    • Base class

    • Task-specific builder

    • Deploy script decorator stack (different from tao-pytorch!)

    • Deploy inference classes

    1. L0 Tests

    • Test directory layout

    • Test markers

    • Trainer dry-run pattern

    • Deploy test pattern (subprocess)

    1. Packaging (setup.py)

Canonical patterns extracted from the TAO submodules. Always read the actual source files before implementing — these are guides, not templates to copy blindly.

1. Config Dataclasses (tao-core)

Location: tao-core/nvidia_tao_core/config/<model_name>/default_config.py

Field types

All fields use typed constructors from nvidia_tao_core.config.utils.types:

from nvidia_tao_core.config.utils.types import (
    BOOL_FIELD, STR_FIELD, INT_FIELD, FLOAT_FIELD,
    LIST_FIELD, DICT_FIELD, DATACLASS_FIELD,
)
TypeUsage
STR_FIELD(value=..., valid_options=..., description=...)String fields, optionally constrained
INT_FIELD(value=..., valid_min=..., valid_max=..., automl_enabled=...)Integers
FLOAT_FIELD(value=..., valid_min=..., math_cond=..., automl_enabled=...)Floats
BOOL_FIELD(value=..., description=...)Booleans
LIST_FIELD(arrList=[...], description=...)Lists
DICT_FIELD({...}, default_value={...}, description=...)Dicts
DATACLASS_FIELD(DataclassInstance())Nested dataclasses
Base classes for sub-configs
from nvidia_tao_core.config.common.common_config import (
    CommonExperimentConfig,  # Top-level base
    TrainConfig,
    EvaluateConfig,
    InferenceConfig,
    ExportConfig,
    GenTrtEngineConfig,
    TrtConfig,
    CalibrationConfig,
)
Top-level ExperimentConfig pattern
@dataclass
class ExperimentConfig(CommonExperimentConfig):
    model:        ModelConfig           = DATACLASS_FIELD(ModelConfig())
    dataset:      DatasetConfig         = DATACLASS_FIELD(DatasetConfig())
    train:        TrainExpConfig        = DATACLASS_FIELD(TrainExpConfig())
    evaluate:     EvalExpConfig         = DATACLASS_FIELD(EvalExpConfig())
    inference:    InferenceExpConfig    = DATACLASS_FIELD(InferenceExpConfig())
    export:       ExportExpConfig       = DATACLASS_FIELD(ExportExpConfig())
    gen_trt_engine: GenTrtEngineExpConfig = DATACLASS_FIELD(GenTrtEngineExpConfig())

    def __post_init__(self):
        if self.model_name is None:
            self.model_name = "<model_name>"
Model parameters mapping

Location: tao-core/nvidia_tao_core/config/<model_name>/model_params_mapping.py

Maps backbone variant names to their output embedding dimensions. Used by build_model() to auto-wire the head's in_channels:

map_params = {
    "head": {
        "in_channels": {
            "vit_base_patch16": 768,
            "vit_large_patch16": 1024,
            # ... one entry per backbone variant
        }
    }
}

# Optional: map input resolutions for backbones that require non-224 sizes
map_input_lr_head = {
    "vit_large_patch14_dinov2_swiglu_legacy": 518,
}

Reference: tao-core/nvidia_tao_core/config/classification_pyt/default_config.py, model_params_mapping.py

2. Backbone (tao-pytorch)

Location: tao-pytorch/nvidia_tao_pytorch/cv/backbone_v2/

Required abstract methods on BackboneBase

New backbones dual-inherit from both the underlying model (HF/timm) AND BackboneBase. The BackboneMeta metaclass automatically calls _post_init() after __init__, which runs freeze_backbone() and set_grad_checkpointing().

from nvidia_tao_pytorch.cv.backbone_v2.backbone_base import BackboneBase
from nvidia_tao_pytorch.cv.backbone_v2 import BACKBONE_REGISTRY

class MyBackbone(SomeHFOrTimmModel, BackboneBase):
    """Dual-inherit: HF/timm model provides the architecture,
    BackboneBase provides TAO integration (freezing, checkpointing, registry)."""

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)

    # --- 6 abstract methods (all required) ---
    def get_stage_dict(self) -> dict:
        """Map stage-index -> nn.Module for layer freezing."""
        return {0: self.patch_embed, 1: self.blocks[:4], 2: self.blocks[4:8], ...}

    def get_classifier(self) -> nn.Module:
        """Return the classification head (e.g., self.head)."""
        return self.head

    def reset_classifier(self, num_classes, **kwargs):
        """Replace head for different num_classes."""
        self.head = nn.Linear(self.embed_dim, num_classes)

    def forward_pre_logits(self, x):
        """Features WITHOUT head. Returns [B, embed_dim]."""
        ...

    def forward_feature_pyramid(self, x, indices=None, **kwargs):
        """Multi-scale features for detection/segmentation.
        Returns dict {scale: feature_tensor}. Classification can return {0: features}."""
        ...

    def forward(self, x):
        """Full forward pass with head. Returns logits."""
        return self.get_classifier()(self.forward_pre_logits(x))

# --- Factory functions (one per variant) ---
@BACKBONE_REGISTRY.register()
def my_backbone_base(**kwargs):
    """kwargs from build_model(): num_classes, freeze_at, freeze_norm, export, img_size, ..."""
    return MyBackbone(embed_dim=768, depth=12, num_heads=12, **kwargs)

@BACKBONE_REGISTRY.register()
def my_backbone_large(**kwargs):
    return MyBackbone(embed_dim=1024, depth=24, num_heads=16, **kwargs)

Key pattern from existing backbones (e.g., vit.py):

  • VisionTransformer(TimmVisionTransformer, BackboneBase) — wraps timm's ViT

  • Overrides forward_pre_logits to handle positional encoding interpolation

  • Each variant (vit_base_patch16, vit_large_patch16, etc.) is a factory function with fixed architecture params

Already-registered backbones in backbone_v2/

vit.py, swin.py, resnet.py, convnext.py, convnext_v2.py, dino_v2.py, fan.py, fastervit.py, gcvit.py, hiera.py, mit.py, edgenext.py, efficientvit.py, radio.py, siglip2.py, open_clip.py

Reference: tao-pytorch/nvidia_tao_pytorch/cv/backbone_v2/backbone_base.py

3. build_model() Pattern (tao-pytorch)

Location: tao-pytorch/nvidia_tao_pytorch/cv/<model_name>/model/classifier.py (or equivalent)

The build_model() function is the core integration point between config, backbone registry, and pretrained weights:

from nvidia_tao_pytorch.cv.backbone_v2.registry import BACKBONE_REGISTRY

def build_model(experiment_config, export=False):
    model_config = experiment_config.model
    backbone_type = model_config.backbone.type

    # 1. Instantiate backbone from registry
    model = BACKBONE_REGISTRY.get(backbone_type)(
        num_classes=experiment_config.dataset.num_classes,
        freeze_at='all' if model_config.backbone.freeze_backbone else None,
        freeze_norm=model_config.backbone.freeze_norm,
        export=export
    )

    # 2. Ensure head remains trainable even if backbone is frozen
    if model_config.backbone.freeze_backbone:
        head = model.get_classifier()
        for p in head.parameters():
            p.requires_grad = True
        head.train()

    # 3. Load pretrained weights with adaptation
    if model_config.backbone.pretrained_backbone_path:
        state_dict = load_pretrained_weights(
            model_config.backbone.pretrained_backbone_path,
            parser=cls_parser,       # Strips "module." prefix from DDP checkpoints
            ptm_adapter=ptm_adapter  # Maps prefixes from other TAO model types
        )

        # Special handling: ViT position embedding interpolation
        if isinstance(model, DINOV2):
            state_dict = interpolate_vit_checkpoint(state_dict, ...)

        msg = model.load_state_dict(state_dict, strict=False)
        logger.info(f"Loaded: {msg}")

    return model
StateDictAdapter for cross-model checkpoint loading
from nvidia_tao_pytorch.cv.classification_pyt.model.utils import StateDictAdapter

ptm_adapter = StateDictAdapter()
ptm_adapter.add("mae", "model.encoder.")           # MAE checkpoints
ptm_adapter.add("classification", "model.")         # Classification checkpoints
ptm_adapter.add("rtdetr", "model.model.backbone.")  # RT-DETR checkpoints

Reference: tao-pytorch/nvidia_tao_pytorch/cv/classification_pyt/model/classifier.py, model/utils.py

4. PyTorch Lightning Module (tao-pytorch)

Base class: nvidia_tao_pytorch.core.lightning.tao_lightning_module.TAOLightningModule

TAOLightningModule provides

  • self.experiment_spec — stored config

  • configure_callbacks() — default TAOStatusLogger, ModelCheckpoint, TAOExceptionCheckpoint

  • _dataloader_batch_check() — validates dataset_size >= total_batch_size

  • on_fit_start(), on_validation_start(), on_test_start(), on_predict_start() — auto-validation

  • on_load_checkpoint() — handles encrypted checkpoint decryption

Constructor pattern
class <ModelName>PlModel(TAOLightningModule):
    def __init__(self, experiment_spec, export=False):
        super().__init__(experiment_spec)
        self.checkpoint_filename = "<model_name>_model"  # MUST set
        self.dataset_config  = self.experiment_spec.dataset
        self.model_config    = self.experiment_spec.model
        self.train_config    = self.experiment_spec.train
        self.eval_config     = self.experiment_spec.evaluate
        self.infer_config    = self.experiment_spec.inference
        self._build_model(export)
        self._build_criterion()
Callbacks (configure_callbacks)
from pytorch_lightning.callbacks import ModelCheckpoint, LearningRateMonitor
from nvidia_tao_pytorch.core.callbacks.loggers import TAOStatusLogger
from nvidia_tao_pytorch.core.callbacks.ema import EMA, EMAModelCheckpoint

# Always include:
callbacks = [TAOStatusLogger(results_dir, append=True), lr_monitor]
# Checkpoint convention:
ModelCheckpoint.FILE_EXTENSION = ".pth"
ModelCheckpoint.CHECKPOINT_EQUALS_CHAR = "_"
ModelCheckpoint.CHECKPOINT_NAME_LAST = f"{self.checkpoint_filename}_latest"
Status logging
import nvidia_tao_pytorch.core.loggers.api_logging as status_logging

status_logging.get_status_logger().kpi = {"train_loss": ..., "val_acc": ...}
status_logging.get_status_logger().write(
    message="...", status_level=status_logging.Status.RUNNING
)
Checkpoint save identifier
def on_save_checkpoint(self, checkpoint):
    checkpoint["tao_model"] = "<model_name>"

Reference: tao-pytorch/nvidia_tao_pytorch/core/lightning/tao_lightning_module.py, cv/classification_pyt/model/classifier_pl_model.py

5. Script Entrypoints (tao-pytorch scripts)

Decorator stack (all scripts use this)
from nvidia_tao_pytorch.core.hydra.hydra_runner import hydra_runner
from nvidia_tao_pytorch.core.decorators.workflow import monitor_status
from nvidia_tao_pytorch.core.tlt_logging import obfuscate_logs
from nvidia_tao_core.config.<model_name>.default_config import ExperimentConfig

spec_root = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))

@hydra_runner(
    config_path=os.path.join(spec_root, "experiment_specs"),
    config_name="experiment_spec",
    schema=ExperimentConfig
)
@monitor_status(name="<ModelName>", mode="<train|evaluate|inference|export>")
def main(cfg: ExperimentConfig) -> None:
    obfuscate_logs(cfg)
    ...
Train script (initialize_train_experiment)
from nvidia_tao_pytorch.core.initialize_experiments import initialize_train_experiment

def run_experiment(experiment_config, key, lightning_module):
    resume_ckpt, trainer_kwargs = initialize_train_experiment(experiment_config, key)
    dm = <ModelName>DataModule(experiment_config.dataset)
    dm.setup(stage="fit")
    model = lightning_module(experiment_config)
    trainer = Trainer(**trainer_kwargs,
                      gradient_clip_val=experiment_config.train.clip_grad_norm)
    trainer.fit(model, dm, ckpt_path=resume_ckpt)

initialize_train_experiment() handles: results_dir creation, GPU config, checkpoint resolution, distributed strategy setup, logger initialization.

CLI Entrypoint (model-level CLI)
from nvidia_tao_pytorch.core.entrypoint import get_subtasks, launch, command_line_parser

def main():
    parser = argparse.ArgumentParser("<model_name>", ...)
    subtasks = get_subtask_list()
    args, unknown_args = command_line_parser(parser, subtasks)
    launch(vars(args), unknown_args, subtasks, network="<model_name>")

get_subtasks(scripts) auto-discovers all .py files in the scripts/ package. launch() constructs python <script.py> --config-path ... --config-name ... and runs it as a subprocess with GPU configuration.

Reference: tao-pytorch/nvidia_tao_pytorch/core/entrypoint.py, cv/classification_pyt/entrypoint/classification.py, cv/classification_pyt/scripts/train.py

6. ONNX Export (tao-pytorch)

from nvidia_tao_pytorch.cv.classification_pyt.utils.onnx_export import ONNXExporter
from nvidia_tao_pytorch.core.utilities import encrypt_onnx

onnx_export = ONNXExporter()
onnx_export.export_model(
    model, batch_size, output_file, dummy_input,
    input_names=["input"], output_names=["output"],
    opset_version=cfg.export.opset_version,
    do_constant_folding=True,
)
onnx_export.check_onnx(output_file)

# Encrypt if needed
if output_file.endswith(".etlt") and key:
    encrypt_onnx(tmp_file_name=tmp_onnx_file, output_file_name=output_file, key=key)

Always use dynamic_axes={"input": {0: "batch"}, "output": {0: "batch"}} for variable batch size.

Reference: tao-pytorch/nvidia_tao_pytorch/cv/classification_pyt/scripts/export.py

7. TensorRT Engine Builder (tao-deploy)

Base class

Location: tao-deploy/nvidia_tao_deploy/engine/builder.py

from nvidia_tao_deploy.engine.builder import EngineBuilder  # Abstract base class

class EngineBuilder(ABC):
    def __init__(self, batch_size, verbose, max_batch_size, opt_batch_size,
                 min_batch_size, workspace, strict_type_constraints, force_ptq,
                 is_qat, timing_cache_path, strongly_typed):
        self.builder = trt.Builder(self.trt_logger)
        self.config = self.builder.create_builder_config()
        ...
Task-specific builder

For classification: reuse ClassificationEngineBuilder from tao-deploy/nvidia_tao_deploy/cv/classification_tf1/engine_builder.py.

For other tasks: subclass EngineBuilder directly, implementing:

  • set_input_output_node_names()

  • Any task-specific preprocessing (mean/std, data format)

Deploy script decorator stack (different from tao-pytorch!)
from nvidia_tao_deploy.cv.common.hydra.hydra_runner import hydra_runner    # deploy version
from nvidia_tao_deploy.cv.common.decorators import monitor_status           # deploy version
from nvidia_tao_deploy.cv.common.initialize_experiments import initialize_gen_trt_engine_experiment
from nvidia_tao_deploy.utils.decoding import decode_model
from nvidia_tao_deploy.cv.common.utils import is_qdq_quantized_onnx

The deploy @monitor_status handles: results_dir creation, experiment.yaml save, status.json lifecycle, exception categorization (config errors, validation errors, filesystem errors).

Deploy inference classes
from nvidia_tao_deploy.cv.classification_tf1.inferencer import ClassificationInferencer
from nvidia_tao_deploy.cv.classification_tf1.dataloader import ClassificationLoader

  • ClassificationInferencer — wraps TRT engine, handles infer(imgs) calls

  • ClassificationLoader — NumPy-based batch loader (no PyTorch dependency)

For non-classification tasks, find equivalent inferencer/loader in the task-specific directory.

Reference: tao-deploy/nvidia_tao_deploy/engine/builder.py, cv/classification_pyt/scripts/gen_trt_engine.py, cv/classification_pyt/scripts/inference.py

8. L0 Tests

Test directory layout
tao-pytorch/tests/cv_unit_test/<model_name>/
├── conftest.py           # Fixtures: _train_spec, _test_dir, etc.
├── test_model.py         # build_model() with various backbones
├── test_trainer.py       # PL Trainer fit/evaluate/inference dry-runs
├── test_dataloader.py    # Data loading pipeline
├── test_config.py        # Config loading & schema validation
└── test_export.py        # ONNX export

tao-deploy/tests/<model_name>/
└── test_<model_name>.py  # gen_trt_engine, inference, evaluate
Test markers
@pytest.mark.cv_unit
@pytest.mark.<model_name>
@pytest.mark.train          # or @pytest.mark.evaluate, .inference
Trainer dry-run pattern
@pytest.mark.parametrize("backbone", TEST_TOPOLOGIES)
def test_trainer_fit(_test_dir, _train_spec, backbone):
    _train_spec.model.backbone.type = backbone
    dm = <ModelName>DataModule(_train_spec.dataset)
    dm.setup(stage="fit")
    model = <ModelName>PlModel(_train_spec)
    trainer = Trainer(devices=_train_spec.train.num_gpus,
                      default_root_dir=_train_spec.results_dir,
                      accelerator='gpu', fast_dev_run=True)
    trainer.fit(model, dm)
Deploy test pattern (subprocess)
def test_gen_trt_engine(model_path, spec_path, tmp_path):
    cmd = f"python {gen_trt_engine_script} -e {spec_path} ..."
    result = subprocess.run(cmd, shell=True, capture_output=True)
    assert result.returncode == 0
    assert (tmp_path / "model.engine").exists()

Reference: tao-pytorch/tests/cv_unit_test/classification_pyt/test_trainer.py, tao-deploy/tests/

9. Packaging (setup.py)

tao-pytorch console_scripts
# In tao-pytorch/setup.py, entry_points.console_scripts:
'<model_name>=nvidia_tao_pytorch.cv.<model_name>.entrypoint.<model_name>:main',
tao-deploy console_scripts
# In tao-deploy/setup.py, entry_points.console_scripts:
'<model_name>=nvidia_tao_deploy.cv.<model_name>.entrypoint.<model_name>:main',

Reference: tao-pytorch/setup.py, tao-deploy/setup.py

10. Core Utilities Summary

UtilityImportPurpose
obfuscate_logs(cfg)nvidia_tao_pytorch.core.tlt_loggingHide encryption keys in logs
expand_path(path)nvidia_tao_pytorch.core.path_utilsSafe tilde expansion + absolute path
get_global_rank()nvidia_tao_pytorch.core.distributed.commDDP rank (0 if not distributed)
get_world_size()nvidia_tao_pytorch.core.distributed.commNumber of processes
is_master_node()nvidia_tao_core.distributed.utilsMulti-framework master check
get_latest_checkpoint()nvidia_tao_pytorch.core.utilitiesFind latest .pth in results_dir
TLTPyTorchCookbooknvidia_tao_pytorch.core.cookbooksEncryption key management

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs Python, Needs Node.js

The author says: Requires Python 3.10+, NVIDIA driver, CUDA 13.0+, docker + nvidia-container-toolkit, an NGC API key (`docker login nvcr.io`), an HF_TOKEN, and access to the TAO Toolkit container images on `nvcr.io` for `tao-pytorch`, `tao-deploy`, and (optionally) `tao-dataservices` — Phase 0 asks the user for the exact image references and prepares them locally as `tao-pytorch-base:latest`, `tao-deploy-base:latest`, `tao-dataservices-base:latest`. Local clones of `tao-core`, `tao-pytorch`, `tao-deploy`,...

Security

What our scanner noticed· Medium risk

  • Large base64-encoded blobs
  • Instructions that use the network

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by
NVIDIA Corporation
Publisher
NVIDIA (Official)
Source
NVIDIA/skills · skills/tao-port-huggingface-model · commit fce3cca
Licence
Apache License 2.0 (redistribution allowed)
  • LICENSE (added by AutoGineer from LICENSE-APACHE)
  • ATTRIBUTION.md (written by AutoGineer)
Synced

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-port-huggingface-model.zip