Tao Port Huggingface Model

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Workflow
Software development
Apache-2.0
Medium risk

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

Download the skill as a ZIP:
https://autogineer.com/agent-skills/nvidia/tao-port-huggingface-model/download

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).

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

Phase	Goal	Reference
0	Prerequisites + TAO Toolkit images + local image tags	phase-0-prereqs.md
1	Inputs, HF-inspection container, validate model + dataset	phase-1-inspection.md, hf-inspection.md
2	Closest existing TAO reference model	phase-2-codebase.md, task-type-guide.md
3	tao-core config + tao-pytorch trainer / eval / inference	phase-3-implementation.md, tao-patterns.md, repo-structure.md
4	ONNX export + tao-deploy TRT engine / inference / eval	phase-4-deploy.md
5	Packaging (`console_scripts`) + L0 tests	phase-5-packaging.md
6	Container testing + end-to-end validation	phase-6-container-tests.md, docker-patterns.md
7	(conditional) Accuracy / latency / size tuning	phase-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

ATTRIBUTION.mdAdded by AutoGineer948 B
BENCHMARK.md7.0 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md17.7 KiB
config/skillspector-baseline.yaml2.6 KiB
eval.config3.2 KiB
evals/evals.json809 B
references/cross-cutting.md10.1 KiB
references/docker-patterns.md15.9 KiB
references/hf-inspection.md6.2 KiB
references/phase-0-prereqs.md9.9 KiB
references/phase-1-inspection.md10.1 KiB
references/phase-2-codebase.md7.3 KiB
references/phase-3-implementation.md34.7 KiB
references/phase-4-deploy.md12.7 KiB
references/phase-5-packaging.md3.3 KiB
references/phase-6-container-tests.md15.4 KiB
references/phase-7-optimization.md7.8 KiB
references/repo-structure.md10.6 KiB
references/tao-patterns.md20.1 KiB
references/task-type-guide.md16.4 KiB
references/workflow-consistency.md27.5 KiB
skill-card.md4.8 KiB
skill.oms.sig8.5 KiB

references/phase-1-inspection.md

Full Phase 1 walkthrough for the tao-port-huggingface-model skill — credential gathering, branch creation, launching the long-lived tao-hf-inspect Docker container (no host venv), model/dataset inspection via docker exec, and the Phase 1 gate. See hf-inspection.md for a generic HF-inspection cheat sheet.

Phase 1 — Information Gathering & Validation

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.

Phase 1 — Information Gathering & Validation
- 1.1 Gather credentials, targets, and locate repos
- 1.2 Create a consistent working branch across all repos
- 1.3 Set up an isolated environment for HF inspection
- 1.4 Validate that the model is a Computer Vision model
- 1.5 Fetch the model architecture and checkpoint
- 1.6 Verify ONNX exportability
- 1.7 Clean up Phase 1 environment
- Phase 1 Gate — Confirm before proceeding:

1.1 Gather credentials, targets, and locate repos

Ask the user for:

HuggingFace Model ID — e.g., google/vit-base-patch16-224
HuggingFace Access Token (HF_TOKEN) — required for gated models
Model short-name for TAO — a snake_case identifier used for directory names and class names (e.g., vit_base_p16)
Do you already have the TAO repos cloned locally? Ask for the paths to tao-core, tao-pytorch, tao-deploy, and tao-dataservices. If the user provides paths, verify they exist and use them. Only clone repos that are missing.

If any repos need to be cloned, ask the user where they'd like them cloned to (default: current working directory), then clone only the missing ones:

# Only clone what's needed — skip repos the user already has
git clone <tao-core-url> /path/to/tao-core
git clone <tao-pytorch-url> /path/to/tao-pytorch
git clone <tao-deploy-url> /path/to/tao-deploy
git clone <tao-dataservices-url> /path/to/tao-dataservices

After cloning, each repo (tao-pytorch, tao-deploy, tao-dataservices) will have a tao-core/ submodule inside it. This submodule points to the original commit and should NOT be used — always use our top-level tao-core/ clone instead (see "Submodule Override Strategy" above).

1.2 Create a consistent working branch across all repos

Before any implementation work, create a new branch in every repo so changes are isolated and consistent:

Ask the user for:
- Branch name — e.g., feature/add-vit-base-p16
- Base branch — default is main. Ask if they want a different base.
Create the branch in all repos:

for repo in tao-core tao-pytorch tao-deploy tao-dataservices; do
  cd /path/to/$repo
  git checkout <base_branch>
  git pull origin <base_branch>
  git checkout -b <branch_name>
  cd -
done

Important: This branch is local only — it will NOT be pushed. It just keeps changes organized and makes it easy to diff against the base branch.

1.3 Set up an isolated environment for HF inspection

All Phase 1 Python work runs inside the prepared tao-pytorch-base:latest container (built in Phase 0) — do NOT install into the host Python. That image already ships torch, transformers, onnx, and timm, and is the same image used in Phases 3/4/6, so there is no need to maintain a separate host venv or to apt-install python3-venv / python3-pip on the host.

set -a; source /path/to/.env; set +a   # omit if already exported
# Scratch dir on the host, bind-mounted into the container as /workspace.
# The directory is owned by the host user that created it, and we run the
# container as that same UID/GID via --user below, so no further chmod is
# needed.
mkdir -p ./.phase1/cache

# Launch a long-lived inspection container so each probe step is a quick `docker exec`.
docker rm -f tao-hf-inspect 2>/dev/null || true
docker run -d --name tao-hf-inspect \
  --user $(id -u):$(id -g) \
  -v "$(pwd)/.phase1":/workspace \
  -e HF_HOME=/workspace/cache -e HF_TOKEN \
  -w /workspace \
  tao-pytorch-base:latest sleep infinity

--user $(id -u):$(id -g) keeps any files written under ./.phase1 (HF cache, the ONNX scratch file) owned by the host user. Use --gpus all only if a probe step needs GPU; AutoConfig / AutoModel / ONNX export are CPU-only.

If tao-pytorch-base:latest is unavailable (e.g. Phase 0 was skipped on a machine that only has CPU), fall back to a small CPU-only image. Note the extra HOME=/workspace + PIP_USER=1 env: python:3.12-slim's system site-packages (/usr/local/lib/python3.12/site-packages) is root-owned, so the pip install would fail with PermissionError once --user $(id -u):$(id -g) drops root. Setting HOME + PIP_USER routes the install into /workspace/.local/lib/python3.12/site-packages inside the bind mount, which the host user can write to. Python's site.py then adds that user-site to sys.path automatically for subsequent docker exec probes:

set -a; source /path/to/.env; set +a   # omit if already exported
docker run -d --name tao-hf-inspect \
  --user $(id -u):$(id -g) \
  -v "$(pwd)/.phase1":/workspace \
  -e HOME=/workspace -e PIP_USER=1 \
  -e HF_HOME=/workspace/cache -e HF_TOKEN \
  -e PIP_CACHE_DIR=/workspace/cache/pip \
  -w /workspace \
  python:3.12-slim \
  bash -c "pip install -q transformers huggingface_hub torch onnx timm && sleep infinity"

1.4 Validate that the model is a Computer Vision model

Run the probe via docker exec:

docker exec -e MODEL_ID="$MODEL_ID" tao-hf-inspect python - <<'PY'
import os
from huggingface_hub import model_info
mid = os.environ["MODEL_ID"]; tok = os.environ.get("HF_TOKEN") or None
info = model_info(mid, token=tok)
print(info.pipeline_tag)   # must be: image-classification, object-detection, image-segmentation, etc.
PY

Hard stop: If pipeline_tag is an NLP, audio, or LLM task, halt and inform the user. TAO Toolkit currently supports Computer Vision models only.

1.5 Fetch the model architecture and checkpoint

docker exec -e MODEL_ID="$MODEL_ID" tao-hf-inspect python - <<'PY'
import os
from transformers import AutoModel, AutoConfig
mid = os.environ["MODEL_ID"]; tok = os.environ.get("HF_TOKEN") or None
config = AutoConfig.from_pretrained(mid, token=tok)
model  = AutoModel.from_pretrained(mid, token=tok)
state_dict = model.state_dict()
print(config)
for k, v in list(state_dict.items())[:30]:
    print(k, tuple(v.shape))
PY

Print config to extract: model_type, image_size, hidden_size, num_labels, num_hidden_layers, patch_size
Print the top-level state_dict keys and shapes to understand HF naming conventions
Assess whether the HF task head is separable from the backbone
Draft a key-name remapping plan for the HF-to-TAO state_dict conversion

1.6 Verify ONNX exportability

docker exec -e MODEL_ID="$MODEL_ID" tao-hf-inspect python - <<'PY'
import os, torch
from transformers import AutoConfig, AutoModel
mid = os.environ["MODEL_ID"]; tok = os.environ.get("HF_TOKEN") or None
config = AutoConfig.from_pretrained(mid, token=tok)
model  = AutoModel.from_pretrained(mid, token=tok).eval()
img_size = getattr(config, "image_size", 224)
if isinstance(img_size, int):
    img_size = (img_size, img_size)
dummy = torch.randn(1, 3, *img_size)
torch.onnx.export(model, dummy, "/workspace/tao_hf_test.onnx",
    input_names=["input"], output_names=["output"],
    dynamic_axes={"input": {0: "batch"}, "output": {0: "batch"}},
    opset_version=17)
print("ONNX export OK")
PY

If this fails, identify the problematic ops and apply workarounds before starting TAO integration:

Unsupported op → Replace with ONNX-compatible equivalent (e.g., replace torch.einsum with explicit matmul/permute, replace custom CUDA kernels with pure PyTorch ops)
Dynamic control flow (if/else on tensor values) → Rewrite as static ops or use torch.where()
Unsupported attention variant → Rewrite using standard nn.MultiheadAttention or explicit Q/K/V matmuls
Try higher opset → opset_version=17 or 18 supports more ops than older versions
TensorRT compatibility → After ONNX export succeeds, test with trtexec inside the prepared tao-deploy container (the host does not have TensorRT):
```
docker run --rm --gpus all \
  --user $(id -u):$(id -g) \
  -v "$(pwd)/.phase1":/workspace \
  -w /workspace \
  tao-deploy-base:latest \
  trtexec --onnx=/workspace/tao_hf_test.onnx --buildOnly
```
If TRT fails on specific layers, those ops will need to be rewritten in the TAO implementation — record them now
If export fundamentally cannot work (e.g., architecture uses dynamic shapes that vary per-input), inform the user — the model may not be suitable for TensorRT deployment

1.7 Clean up Phase 1 environment

After all inspection is complete and findings are recorded:

docker rm -f tao-hf-inspect

# Remove the host scratch dir (HF cache + tao_hf_test.onnx + pip cache).
# Keep .phase1 around between reruns if you want to skip the model redownload.
rm -rf ./.phase1

Phase 1 Gate — Confirm before proceeding:

[ ] All 4 TAO repos located or cloned
[ ] Consistent working branch created across all repos
[ ] pipeline_tag is a supported CV task
[ ] model_type, image_size, hidden_size, num_labels extracted
[ ] Top-level state_dict keys documented, remapping plan drafted
[ ] ONNX export sanity check passed (or failure mode understood)
[ ] User confirmed the model short-name and task type

Present findings to the user and get confirmation before proceeding to implementation.

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

2026-11-09

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

From the author

AutoGineer summary

What it does

When to use it

Advantages

Usage examples

Install

Documentation by NVIDIA Corporation

TAO-HF Integration Skill

Local-Only Rule

Submodule Override Strategy

Execution platform

Phase Map

Development Loop

Debugging Playbook

Environment Isolation Strategy

Phase 0 — Prerequisites Check

Phase 1 — Information Gathering & Validation

Phase 2 — Codebase Exploration

Phase 3 — TAO Core Configuration & Native Implementation

Phase 4 — Export, Deployment & TensorRT Integration

Phase 5 — Packaging & L0 Testing

Cross-Phase Data Flow & Consistency Verification

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

Phase 7 — Optimization & Tuning (conditional)

Argument

Files

references/phase-1-inspection.md

Phase 1 — Information Gathering & Validation

Contents

1.1 Gather credentials, targets, and locate repos

1.2 Create a consistent working branch across all repos

1.3 Set up an isolated environment for HF inspection

1.4 Validate that the model is a Computer Vision model

1.5 Fetch the model architecture and checkpoint

1.6 Verify ONNX exportability

1.7 Clean up Phase 1 environment

Phase 1 Gate — Confirm before proceeding:

Requirements and compatibility

Security

Attribution and licence

Related skills

Implement

Source Driven Development

Write Swift

Cardputer Buddy

Typescript Pro

Scala Pro

Download ZIP