i4h-workflow-finetune

par nvidia

Affiner une politique GR00T ou openpi PI0 sur un jeu de données LeRobot. Utiliser lorsqu’on demande d’affiner, d’entraîner ou de post-entraîner une politique sur des démonstrations ; pas pour évaluer une…

npx skills add https://github.com/nvidia/skills --skill i4h-workflow-finetune

i4h Workflow — Finetune

Purpose

Fine-tune a GR00T or openpi PI0 policy on an existing LeRobot dataset. Use when asked to finetune, train, or post-train a policy on recorded demos.

Base Code

These steps drive the i4h-workflows base code (the workflows/agentic/ tree). To reuse an existing checkout, set I4H_WORKFLOWS to its path (no clone happens). Otherwise this resolves the current repo, or clones to ~/i4h-workflows — pick that default without prompting. Run every command below from the resolved root:

# Resolve the i4h-workflows base code (provides workflows/agentic/).
ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"
if [ ! -d "$ROOT/workflows/agentic" ]; then
  ROOT="${I4H_WORKFLOWS:-$HOME/i4h-workflows}"
  [ -d "$ROOT/workflows/agentic" ] || git clone https://github.com/isaac-for-healthcare/i4h-workflows "$ROOT"
fi
export I4H_WORKFLOWS="$ROOT"; cd "$ROOT"

Basics

  • The dataset path must be an existing LeRobot directory with meta/info.json.
  • Train support is determined by policy.train_module in workflows/agentic/config/environments/<env>.yaml. A null value means inference-only.
  • assemble_trocar is inference-only.

Stack Map

EnvStackCLI
scissor_pick_and_placegr00t_n15i4h-agentic-gr00t-n15-train
locomanip_tray_pick_and_placegr00t_n16i4h-agentic-gr00t-n16-train
locomanip_push_cartgr00t_n16i4h-agentic-gr00t-n16-train
ultrasound_liver_scanopenpi_pi0i4h-agentic-openpi-pi0-train

N1.6 locomanip envs share policy.locomanip.train.

Preflight

test -f "${DATASET_PATH}/meta/info.json"
nvidia-smi --query-gpu=name --format=csv,noheader | wc -l
workflows/agentic/policy/<stack>/run.sh --list-envs

Run

Run the steps below in order. Each step is a separate bash call; variables persist in the local agent's tmux session.

Step 1 — setup and resolve dataset

REPO_ROOT="${I4H_WORKFLOWS:-$(git rev-parse --show-toplevel 2>/dev/null)}"; [ -d "$REPO_ROOT/workflows/agentic" ] || REPO_ROOT="$HOME/i4h-workflows"
ENV_ID=scissor_pick_and_place
STACK_DIR=gr00t_n15
TRAIN_CLI=i4h-agentic-gr00t-n15-train
RUNS_ROOT="${REPO_ROOT}/workflows/agentic/runs"

# Point DATASET_PATH at a converted LeRobot dataset dir (absolute; must contain meta/info.json),
# produced by [[i4h-workflow-dataset-convert]]. List candidates:
#   find "${RUNS_ROOT}" "${HF_LEROBOT_HOME:-$HOME/.cache/huggingface/lerobot}" -name info.json -path '*/meta/*' -printf '%h\n' | sed 's#/meta$##' | sort -u
DATASET_PATH="${DATASET_PATH:-}"
if [ ! -f "${DATASET_PATH%/}/meta/info.json" ]; then
  echo "finetune: set DATASET_PATH to a LeRobot dataset dir with meta/info.json (got '${DATASET_PATH:-<unset>}'). Candidates:" >&2
  find "${RUNS_ROOT}" "${HF_LEROBOT_HOME:-$HOME/.cache/huggingface/lerobot}" -name info.json -path '*/meta/*' -printf '%h\n' 2>/dev/null | sed 's#/meta$##' | sort -u | head
  exit 1
fi

RUN_DIR="${RUNS_ROOT}/finetune_${ENV_ID}_$(date +%Y%m%d_%H%M%S)"
OUT="${RUN_DIR}/checkpoint"
export TMPDIR=/tmp   # short path: torch DataLoader FD-sharing socket must fit AF_UNIX's 108-byte limit
mkdir -p "${OUT}" "${RUN_DIR}/logs"
ln -sfn "${RUN_DIR}" "${RUNS_ROOT}/.latest"

Step 2 — train

uv --directory "${REPO_ROOT}/workflows/agentic/policy/${STACK_DIR}" run "${TRAIN_CLI}" \
  --env "${ENV_ID}" \
  --dataset-path "${DATASET_PATH}" \
  --output-dir "${OUT}" \
  --max-steps 1000 \
  --save-steps 1000 \
  --num-gpus 1 \
  2>&1 | tee "${RUN_DIR}/logs/finetune.log"

Tyro flags use kebab case (--max-steps, not --max_steps).

Common Flags

  • --dataset-path PATH (required)
  • --output-dir PATH
  • --base-model-path PATH_OR_REPO overrides YAML policy.model_repo
  • --max-steps N, --save-steps N
  • --batch-size N, --learning-rate FLOAT
  • --no-tune-visual — freeze the vision backbone (trains the action head + projector only): ~2× faster, ~half the memory, less overfitting. Good default for small datasets; unfreeze only with lots of data + a real visual domain gap.
  • --num-gpus N — must not exceed visible GPUs
  • --report-to tensorboard|wandb

Verify

  • Checkpoint directory ${OUT}/checkpoint-<N> contains model-0000*-of-*.safetensors, experiment_cfg/, processor/.
  • Log contains train_loss lines and a final 'train_runtime': ... summary.

Prerequisites

  • Workflow set up via [[i4h-workflow-setup]] (the stack's .venv must exist).
  • An existing LeRobot dataset directory with meta/info.json.
  • A train-capable env: policy.train_module non-null in workflows/agentic/config/environments/<env>.yaml (assemble_trocar is inference-only).
  • At least one visible GPU (--num-gpus must not exceed visible GPUs).

Limitations

  • Inference-only envs (null policy.train_module, e.g. assemble_trocar) cannot be fine-tuned.
  • Requires GPU(s); --num-gpus must not exceed the count from nvidia-smi.
  • N1.6 locomanip envs share policy.locomanip.train.
  • Each env maps to one stack/CLI (see Stack Map); the dataset must match that env.

Troubleshooting

  • Error: train CLI / module import fails - Cause: workflow not set up, stack .venv missing. Fix: run [[i4h-workflow-setup]] first.
  • Error: dataset path rejected / missing meta/info.json - Cause: --dataset-path is not a valid LeRobot directory. Fix: point to a converted LeRobot dataset (see Preflight test -f).
  • Error: env is inference-only / no train support - Cause: policy.train_module is null for that env. Fix: choose a train-capable env from the Stack Map.
  • Error: unrecognized flag like --max_steps - Cause: Tyro flags use kebab case. Fix: use --max-steps form.

Final Response

Report env, stack, dataset path, output checkpoint path, train_loss summary, and blockers.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…