tao-run-on-local-docker

Imported from nvidia/skills (skills/tao-run-on-local-docker/SKILL.md) via skills.sh. Install upstream with npx skills add nvidia/skills --skill tao-run-on-local-docker. Copyright stays with the author (Apache-2.0).

Local Docker

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).

Single-node execution platform that runs TAO jobs as named Docker containers on a Docker daemon. The daemon can be local to the agent host or remote through DOCKER_HOST=ssh://user@host / a Docker context. It is useful for development, debugging, small runs, and workflows where a local coding agent submits jobs to a remote GPU box.

Use local Docker when the data is local to the Docker host or accessible through mounted volumes/cloud credentials. Do not use it for remote cluster scheduling, multi-node training, or jobs that need SLURM queueing.

Use remote Docker when the agent is running on a workstation or laptop but the Docker daemon and GPUs are on another single GPU server. In remote Docker mode, all local filesystem paths in specs are interpreted on the remote Docker host, not on the agent machine.

Preflight

The workflow must verify the host GPU runtime before starting Docker jobs. If the check fails, prompt the user to approve the install, run the printed install command, and rerun the preflight.

# Host GPU runtime: NVIDIA driver 580, CUDA 13.0, NVIDIA Container Toolkit 1.19.0.
SB="${TAO_SKILL_BANK_PATH:-${TAO_SKILL_BANK_ROOT:-$PWD}}"
SETUP_SCRIPT="${SB}/skills/platform/tao-setup-nvidia-gpu-host/scripts/setup-nvidia-gpu-host.sh"

bash "$SETUP_SCRIPT" --backend docker --check-only || {
  echo "MISSING: TAO GPU host runtime is not ready."
  echo "After user approval, run:"
  echo "  bash \"$SETUP_SCRIPT\" --backend docker --install --yes"
  exit 1
}

# Mode 1 — direct docker (no Python). All you need is docker + the GPU runtime.
docker info >/dev/null 2>&1 || { echo "MISSING: docker daemon not reachable. Start Docker."; exit 1; }
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi >/dev/null 2>&1 || {
  echo "MISSING: NVIDIA Container Toolkit not installed/configured. See:"
  echo "  bash \"$SETUP_SCRIPT\" --backend docker --install --yes"
  exit 1
}

# Mode 2 — TAO SDK wrapper. Adds Job handles, S3 I/O wrapping, ActionWorkflow.
# Skip this block if Mode 1 is sufficient for the user's request.
# When Mode 2 is in scope, read `tao-skill-bank:tao-run-platform` for the DockerSDK
# kwarg contract, build_entrypoint, and monitoring patterns.
# nvidia-tao-sdk is on public PyPI; the pin below is stamped from the release manifest.
PIN="nvidia-tao-sdk[docker]==7.1.0rc42"  # versions-key: wheels.tao_sdk_docker
python -c "import tao_sdk" 2>/dev/null || python -m pip install "$PIN"
python -c "import docker" 2>/dev/null || python -m pip install "$PIN"
python -c "import tao_sdk, docker"

# DockerSDK attaches every job container to ${DOCKER_NETWORK:-tao_default}.
# Create the network if it is missing; the operation is local and idempotent.
DOCKER_NETWORK_NAME="${DOCKER_NETWORK:-tao_default}"
docker network inspect "$DOCKER_NETWORK_NAME" >/dev/null 2>&1 || \
  docker network create "$DOCKER_NETWORK_NAME" >/dev/null

If a check fails, the agent prompts the user to authorize the install/fix via Bash before proceeding. Pip-installable Python requirements and Docker network creation above are exceptions: install/create them automatically, then rerun preflight.

Credentials

There are no platform credentials required beyond access to the Docker daemon.

Optional environment:

• DOCKER_HOST: Optional Docker daemon URL. If unset, the SDK uses the Docker Python client's normal environment/default socket resolution. Required for the remote-docker platform option.
• DOCKER_NETWORK: Docker network for job containers. Default is tao_default.
• DOCKER_USERNAME: Registry username. Default is $oauthtoken for NGC.
• NGC_KEY: Used when pulling private images from nvcr.io.
• HOST_SSH_PATH: Mounted into AutoML brain containers when they need SSH keys to monitor remote SLURM child jobs.
• ACCESS_KEY, SECRET_KEY, S3_ENDPOINT_URL, S3_BUCKET_NAME: Optional S3-compatible storage settings for jobs that still read/write cloud storage from a local container.

Launch Preflight

Before generating scripts or starting containers:

1. Verify the Docker daemon is reachable, NVIDIA Container Toolkit is registered as a Docker runtime, GPUs and driver version are reported, and a smoke container can see GPUs before launch. For remote Docker, query GPUs through docker run ... nvidia-smi against the remote daemon; do not use local nvidia-smi from the agent machine.
2. Verify every local/file dataset annotation and media path exists on the Docker host.
3. Classify every bind mount as read-only or writable. Writable mounts must use the Docker host user's numeric UID:GID by default, and the container identity (USER/LOGNAME) plus HOME/framework-cache paths must be set and writable for that identity. Direct Docker must pass them explicitly (see "Non-root container identity"); the SDK prepares them under a writable /results bind (with an isolated /tmp fallback only for forced non-root jobs without one). For remote Docker, resolve the identity on the remote host rather than copying the agent laptop's numeric ids.
4. For s3:// datasets/results, verify ACCESS_KEY and SECRET_KEY are set and the exact paths are readable with aws s3 ls. If aws is missing, report the missing dependency and ask before installing it; rerun preflight after installation.
5. Verify model-specific credentials such as HF_TOKEN before launch.
6. Check current GPU occupancy with nvidia-smi and avoid GPUs already used by other running jobs when the user requested that constraint. Show the selected GPU ids in the launch review.
7. For model/container combinations with known architecture limits, compare host GPU compute capability with the container stack before launch. If the selected image cannot JIT or run kernels for the host architecture, block early and ask for a compatible image or platform.

Use the packaged helper for these checks when possible:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/check_tao_launch_preflight.py \
  --platform local-docker \
  --container-image "<selected-image>" \
  --path train_annotation=/abs/path/to/annotations.json \
  --path train_media=/abs/path/to/media

For a remote Docker daemon, use the remote-docker platform and pass or export DOCKER_HOST. The helper verifies remote GPU/runtime readiness and checks remote-host dataset paths through read-only bind mounts:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/check_tao_launch_preflight.py \
  --platform remote-docker \
  --docker-host ssh://user@gpu-host \
  --container-image "<selected-image>" \
  --gpu-smoke-image ubuntu:22.04 \
  --path train_annotation=/remote/data/train/annotations.json \
  --path train_media=/remote/data/train

The --path values above must exist on the remote Docker host. Do not pass paths that exist only on the local laptop or Codex host.

Resolve the UID:GID of the actual submitting user on the remote Docker host, then pass that identity to the SDK explicitly. Do not reuse the client laptop's UID:GID, and do not infer a container user from stat ownership of a shared output directory: that directory may be root:<shared-group> or owned by another group member.

REMOTE_RESULTS=/remote/results
# Use the same SSH account represented by DOCKER_HOST=ssh://user@gpu-host,
# or obtain these two values from the remote administrator.
REMOTE_UID="$(ssh user@gpu-host id -u)"
REMOTE_GID="$(ssh user@gpu-host id -g)"
case "$REMOTE_UID" in
  ''|*[!0-9]*|0)
    echo "A verified non-root remote submitting UID is required."
    exit 1
    ;;
esac
case "$REMOTE_GID" in
  ''|*[!0-9]*)
    echo "A verified numeric remote submitting GID is required."
    exit 1
    ;;
esac
TAO_DOCKER_CONTAINER_USER="$REMOTE_UID:$REMOTE_GID"
export TAO_DOCKER_CONTAINER_USER

# Prove that this exact identity can create and remove a child in the bind.
docker --host "$DOCKER_HOST" run --rm \
  --user "$TAO_DOCKER_CONTAINER_USER" \
  -v "$REMOTE_RESULTS:/ownership-probe" ubuntu:22.04 \
  sh -c 'p=/ownership-probe/.tao-write-delete-probe-$$; touch "$p" && rm "$p"' || {
  echo "Remote submitting identity cannot write/delete under $REMOTE_RESULTS."
  exit 1
}

Non-root container identity

--user <uid>:<gid> is necessary but not sufficient. TAO images provision non-root accounts only at UID 1000 (ubuntu and taotoolkituser, which collide there), so every other numeric UID runs with no /etc/passwd entry. That makes the failure invisible on a UID-1000 workstation and reproducible everywhere else. getpass.getuser() reads LOGNAME/USER/LNAME/ USERNAME and only then falls back to pwd.getpwuid(), so with none of them set the lookup raises before any TAO code runs:

File "/usr/lib/python3.12/getpass.py", line 169, in getuser
    return pwd.getpwuid(os.getuid())[0]
KeyError: 'getpwuid(): uid not found: 1002'

Torch reaches that call while initializing its inductor cache directory during import, so the container exits 1 at startup. Docker also leaves HOME=/ for an unknown UID, which sends framework caches into image-owned paths.

Every direct-Docker launch that passes --user must therefore also pass the identity and cache environment. These mirror what the SDK injects in docker_handler.py; keep the two lists in sync when either changes.

HOST_UID="$(id -u)"; HOST_GID="$(id -g)"
TAO_HOME=/results/.tao-runtime/home        # must live on a writable mount
mkdir -p "$RESULTS_DIR/.tao-runtime/home"

docker run --rm --gpus all --ipc=host \
  --ulimit memlock=-1 --ulimit stack=67108864 \
  --user "$HOST_UID:$HOST_GID" \
  -e USER="$HOST_UID" -e LOGNAME="$HOST_UID" \
  -e HOME="$TAO_HOME" \
  -e XDG_CACHE_HOME="$TAO_HOME/.cache" \
  -e HF_HOME="$TAO_HOME/.cache/huggingface" \
  -e TORCH_HOME="$TAO_HOME/.cache/torch" \
  -e TRITON_CACHE_DIR="$TAO_HOME/.cache/triton" \
  -e TORCHINDUCTOR_CACHE_DIR="$TAO_HOME/.cache/torchinductor" \
  -e MPLCONFIGDIR="$TAO_HOME/.cache/matplotlib" \
  -v "$DATA_DIR:/data:ro" -v "$RESULTS_DIR:/results" -v "$SPECS_DIR:/specs:ro" \
  "$IMAGE" <action> train -e /specs/<spec>.yaml

Numeric USER/LOGNAME values are deliberate: they describe an identity that genuinely has no passwd entry, and they are only consumed for cache-path naming. Do not drop --user to work around a startup getpwuid failure — that trades a startup error for root-owned outputs, which is the more expensive failure to repair.

Multi-GPU and multi-node

Multi-node is not supported on local Docker. One job runs on the local Docker daemon's host with no cross-host coordination.

Multi-GPU on the local host is supported via the NVIDIA Container Toolkit's --gpus flag (--gpus all or --gpus '"device=0,1,2,3"'). DockerSDK.create_job(gpu_count=N) plumbs through to --gpus. Single-host distributed init uses localhost; torchrun --nproc-per-node=N or PyTorch DDP work as usual.

Backend Details

Use the SDK backend value local-docker. The local backend schema has no extra backend details, so most routing is controlled by environment and job parameters:

{
  "backend_type": "local-docker",
  "num_gpu": 1
}

Following the Brev SDK design, platform/control-plane values stay in SDK state and Docker labels. The SDK does not inject BACKEND, HOST_PLATFORM, MONGOSECRET, DOCKER_HOST, or DOCKER_NETWORK into the training container.

Container Execution

The TAO SDK local Docker handler starts containers through the Docker Python client:

• Backend job name uses the tao-job-<job_id> form used by SDK handlers.
• Command is usually ["/bin/bash", "-c", "<job command>"].
• Containers run detached. The SDK keeps containers by default so status and logs remain inspectable, unless DOCKER_AUTO_REMOVE=true.
• With run_as_user=None (the default), the SDK maps a local job to the invoking UID:GID when it has an absolute writable /results bind, preserves local supplementary groups, and prepares HOME/framework caches under /results/.tao-runtime/home. run_as_user=True opts other local mount layouts into user mapping. If the SDK process itself is root, automatic mapping fails closed instead of mapping 0:0; provide the verified submitting non-root UID:GID through container_user. container_user is also the explicit non-root Docker user override for remote hosts. run_as_user=False is the deliberate opt-out for an image proven to require root.
• /dev/shm is mounted as tmpfs.
• The configured Docker network is applied by the Docker daemon for the job container; it is not passed through as a process environment variable.
• Existing containers with the same job id are stopped and removed before a replacement starts.

For GPU access, the handler auto-detects the host type:

• Tegra or Jetson hosts use runtime="nvidia" plus NVIDIA_VISIBLE_DEVICES and NVIDIA_DRIVER_CAPABILITIES=all.
• Standard x86 hosts use Docker device_requests with GPU capabilities.

If num_gpus is 0, no GPUs are assigned. If num_gpus is -1, all visible GPUs are requested. Prefer explicit GPU counts for shared development machines. When explicit device ids are available, prefer them over count-only selection on shared machines so the launch does not steal GPUs occupied by other tasks.

Storage

Local Docker accepts local and file:// paths because the container runs on the same Docker host. Make sure every path in the spec is either:

• mounted into the container by the handler or surrounding service,
• reachable from inside the container already, or
• a cloud URI with matching credentials.

For bind-mounted outputs, host-user ownership is a launch invariant, not a permission-error workaround. Root containers commonly create checkpoint subdirectories as root:root mode 0755; the host user then cannot delete files inside them even if the top-level output directory was pre-created. Container auto-removal also leaves bind-mounted outputs untouched.

Only opt out of host-user mapping (run_as_user=False) when the selected image demonstrably requires root. Record that exception in the launch review, isolate its writable mounts, and normalize every output/cache mount back to the Docker host UID:GID after all terminal exits and cancellations. For remote Docker, pass the remote host's verified non-root identity through container_user; never infer it from the client machine or output-directory owner. Do not begin another experiment until ownership normalization succeeds. If the agent lacks permission to perform or verify that repair, the root-required image cannot be launched on local Docker.

AutoML's default checkpoint retention is stricter: its preflight rejects run_as_user=False, named volumes, remote bind mounts, or an incompatible explicit container_user before launching a trial, because the SDK cannot guarantee host-side deletion. Use those routes for AutoML only when retention is explicitly disabled and an external operator owns artifact cleanup.

For remote/shared filesystems, prefer the platform that owns that filesystem. For example, use SLURM plus lustre:///... for Lustre paths on a cluster.

Monitoring

• The SDK handler maps Docker container state directly: created -> Pending, running/restarting -> Running, paused -> Paused, exit code 0 -> Complete, nonzero exit -> Error.
• Logs come directly from the named container through the Docker Python client (docker logs tao-job-<job_id>).

If the container has exited, died, is being removed, or cannot be found, status reconciliation treats the backend process as terminated.

Cancellation

Cancellation stops the named container. GPU ownership is managed by Docker / the NVIDIA runtime, not by TAO Core's local GPU manager.

Optional: via the TAO SDK

If you want Job handles, S3 I/O wrapping via the SDK's script_runner, or durability across sessions:

import os

from tao_sdk.platforms.docker import DockerSDK

docker_host = os.environ.get('DOCKER_HOST', '')
is_remote = bool(docker_host) and not docker_host.startswith(('unix://', 'npipe://', '/'))
container_user = os.environ.get('TAO_DOCKER_CONTAINER_USER')
if is_remote and not container_user:
    raise RuntimeError('Set TAO_DOCKER_CONTAINER_USER to the remote output owner UID:GID')

sdk = DockerSDK()  # reads DOCKER_HOST, NGC_KEY, S3 creds from env
job = sdk.create_job(
    image='nvcr.io/nvidia/tao/tao-toolkit:7.1.0-pyt',  # versions-key: images.tao_toolkit.pyt
    command='dino train -e /data/spec.yaml',
    gpu_count=1,
    mounts=[
        {'host_path': '/host/data', 'container_path': '/data', 'read_only': True},
        {'host_path': '/host/results', 'container_path': '/results'},
    ],
    container_user=container_user,
)

status = sdk.get_job_status(job.id)
logs = sdk.get_job_logs(job.id, tail=200)

This wraps the same docker run invocation under a Job handle. For S3 I/O, call build_entrypoint(...) first and pass its command so script_runner can perform the declared downloads/uploads. If you do not need job tracking or that wrapper, use docker run directly — no SDK install required.

Failure Modes

Docker client not initialized: Verify the Docker Python package is installed, set DOCKER_HOST if you are not using the default local socket, and confirm the process can talk to the daemon.

GPU assignment failed: Requested GPUs are unavailable, the NVIDIA Container Toolkit is not configured, or the Docker daemon cannot create GPU device requests. Use fewer GPUs, wait for another job to finish, or verify docker run --gpus ... works on the host.

Image pull auth failed: Set a valid NGC_KEY for private nvcr.io images or run docker login nvcr.io -u '$oauthtoken' on the Docker host.

Container exited unexpectedly: Check docker logs tao-job-<job_id>, the configured DOCKER_NETWORK, and the command produced by the SDK action runner.

KeyError: getpwuid(): uid not found: The launch passed --user with a UID that has no /etc/passwd entry in the image and did not pass USER/LOGNAME. Add the identity and cache environment from "Non-root container identity"; do not fall back to running as root.

Path missing inside container: A local path on the host is not necessarily mounted into the job container. Use a path convention supported by the action runner or configure an explicit volume through the surrounding service.

Root-owned bind-mounted results: Stop launching new experiments, identify every writable mount from docker inspect, and have the host administrator repair existing ownership once. Future launches must use host UID:GID mapping and writable HOME/cache redirects. docker rm and DOCKER_AUTO_REMOVE do not repair or delete bind-mounted files.