Tao Run On Docker

Docker for NVIDIA GPU Workloads

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).

The Docker execution platform: a consumer that runs a model/data skill's spec-bundle by implementing four verbs (submit/status/logs/cancel) over the docker CLI, on a local daemon or a remote GPU box via DOCKER_HOST=ssh://. The verbs (§ Execution) sit on top of the docker conventions in the rest of this file — GPU flags, mounts, NGC auth, inspection, error modes — which are the how the model/data skill defers to. Single-node only; for multi-node use SLURM or Kubernetes.

Sources: official Docker CLI reference (https://docs.docker.com/reference/cli/docker/) and NVIDIA Container Toolkit docs.

Prerequisites

1. Host GPU runtime — by default, NVIDIA driver >=580, CUDA Toolkit >=13.0, and NVIDIA Container Toolkit >=1.19.0. If the selected model's references/skill_info.yaml declares runtime_requirements.gpu_host, pass those values to tao-setup-nvidia-gpu-host instead. Model requirements override the defaults for that workflow.
2. Docker — docker --version must return ≥ 20.10. Install: https://docs.docker.com/engine/install/.
3. NGC API key for nvcr.io/* pulls. Get from https://ngc.nvidia.com/.

set -a; source /path/to/.env; set +a   # omit if already exported
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 "$SETUP_SCRIPT" --backend docker --check-only || {
  echo "MISSING: TAO GPU host runtime is not ready."
  echo "After user approval, run (append --yes for non-interactive agent runs):"
  echo "  bash \"$SETUP_SCRIPT\" --backend docker --install"
  exit 1
}

docker --version
docker run --rm --runtime=nvidia --gpus all ubuntu nvidia-smi
[ -n "$NGC_KEY" ] || echo "NGC_KEY unset — cannot pull nvcr.io images"

If the selected model declares runtime_requirements.gpu_host, append the corresponding --min-driver-version, --min-cuda-version, and --min-container-toolkit-version values to both the check and any approved install command. Do not apply one model's override to unrelated workflows.

NGC authentication

set -a; source /path/to/.env; set +a   # omit if already exported
echo "$NGC_KEY" | docker login nvcr.io -u '$oauthtoken' --password-stdin

Persists in ~/.docker/config.json across reboots. Re-run on unauthorized errors.

Execution — the four verbs

Run a spec-bundle by implementing exactly these four verbs, mutating only the job-record. Status values are the fixed vocabulary from tao-artifacts (PENDING RUNNING COMPLETE ERROR CANCELED UNKNOWN); native docker states map below, with the raw state carried in the transition message. $BANK = ${TAO_SKILL_BANK_PATH}.

submit

1. Stage inputs via tao-data-io: it picks the storage tier and returns the mount args + compute-frame paths. Docker uses tier A (bind-mount a host dir, -v /host/data:/data) as the norm, or tier C (pass S3 creds, the container fetches). Author the spec file at <stage>/spec.yaml with those compute-frame paths.
2. Lint the assembled command — redact_secrets.py lint must pass (no inline secrets; pass creds as -e VAR with no value).
3. Open the record — this mints the id and binds results_dir BEFORE launch:

   JOB_ID=$("$BANK/scripts/tao_job_record.py" open \
     --platform docker --image "$IMAGE" \
     --network-arch "$ARCH" --action "$ACTION" \
     --storage-tier "$TIER" --results-root "$RESULTS_ROOT")
   

4. Launch detached, naming the container after the id so the other verbs find it (keep --rm OFF so an exited container stays inspectable):

   set -a; source /path/to/.env; set +a   # omit if already exported
   CID=$(docker run -d --name "$JOB_ID" --label "tao-job=$JOB_ID" \
     --gpus "$GPUS" --ipc=host \
     -v "$STAGE:/workspace" \
     -e AWS_ACCESS_KEY_ID -e AWS_SECRET_ACCESS_KEY -e HF_TOKEN -e NGC_KEY \
     "$IMAGE" <bundle command, reading /workspace/spec.yaml>)
   

5. Record RUNNING: "$BANK/scripts/tao_job_record.py" mark "$JOB_ID" --state RUNNING --backend-ref "$CID".

A submit that skipped step 3 has no id, so it cannot launch — that is the record-then-launch invariant.

status

read -r st code < <(docker inspect --format '{{.State.Status}} {{.State.ExitCode}}' "$JOB_ID" 2>/dev/null) || st=missing

On a terminal state, mark it — and for tier C, tao-data-io uploads results before you docker rm (the container is the only copy).

logs

docker logs --tail "${N:-200}" "$JOB_ID"    # add -f to follow in-turn

cancel

docker rm -f "$JOB_ID"
"$BANK/scripts/tao_job_record.py" mark "$JOB_ID" --state CANCELED --source agent

Local vs remote (DOCKER_HOST)

There is no separate "remote docker" — point the daemon at an SSH-reachable box: export DOCKER_HOST=ssh://user@gpu-host. Every verb above is byte-identical; the docker CLI marshals the request over SSH (reuses your key, avoids nested-quoting the command). One consequence: -v bind-mount sources then refer to the remote host's filesystem, not the launcher — stage data there (tier A) or fetch in-container (tier C). Fallback for sudo docker-only hosts: ssh host 'sudo docker …' (same skill, different prefix).

docker run — canonical flags

set -a; source /path/to/.env; set +a   # omit if already exported
HOST_RESULTS=/host/results
HOST_UID="$(id -u)"
HOST_GID="$(id -g)"
HOST_USER_NAME="$(id -un)"
[ "$HOST_UID" -ne 0 ] || { echo "Refusing writable Docker launch as UID 0" >&2; exit 1; }
HOST_IDENTITY_ARGS=(--user "$HOST_UID:$HOST_GID")
for group_id in $(id -G); do
  [ "$group_id" = "$HOST_GID" ] || HOST_IDENTITY_ARGS+=(--group-add "$group_id")
done
mkdir -p "$HOST_RESULTS/.tao-runtime/home/.cache"/{huggingface,torch,triton,torchinductor,matplotlib}

docker run \
  --gpus all \
  --rm \
  --shm-size=8g \
  "${HOST_IDENTITY_ARGS[@]}" \
  -v /host/data:/data \
  -v "$HOST_RESULTS:/results" \
  -e HOME=/results/.tao-runtime/home \
  -e USER="$HOST_USER_NAME" -e LOGNAME="$HOST_USER_NAME" \
  -e XDG_CACHE_HOME=/results/.tao-runtime/home/.cache \
  -e HF_HOME=/results/.tao-runtime/home/.cache/huggingface \
  -e TORCH_HOME=/results/.tao-runtime/home/.cache/torch \
  -e TRITON_CACHE_DIR=/results/.tao-runtime/home/.cache/triton \
  -e TORCHINDUCTOR_CACHE_DIR=/results/.tao-runtime/home/.cache/torchinductor \
  -e MPLCONFIGDIR=/results/.tao-runtime/home/.cache/matplotlib \
  -e HF_TOKEN -e NGC_KEY \
  <image> \
  <command>

Notes:

• --gpus '"device=0,1"' — select GPUs by id, not by count, on any shared host (double-quote-escaped). A count-based request resolves to the first N devices, so --gpus 1 can only ever land on GPU 0: if GPU 0 is busy, every job OOMs there while the other GPUs sit idle, and there is no way to steer it — -e NVIDIA_VISIBLE_DEVICES is overwritten by --gpus. Read current occupancy (nvidia-smi --query-gpu=index,memory.used --format=csv) and pass the free ids. Ids may also be GPU UUIDs. Without nvidia-container-toolkit: could not select device driver "" with capabilities: [[gpu]].
• --rm — clean up the container at exit; omit when you want docker logs after exit.
• --shm-size=8g — torchrun + PyTorch DataLoaders exhaust the default 64 MB /dev/shm otherwise; size it for multi-GPU training and raise (e.g. 16g) if you still hit Bus error.
• --user "$(id -u):$(id -g)" — required by default whenever a bind mount is writable. It prevents root-owned checkpoint trees that the submitting host user cannot clean up.
• Refuse UID 0 for the canonical writable-bind path. If the launcher itself is root, obtain the verified non-root submitting UID:GID explicitly; never infer it from the output-directory owner.
• --group-add <gid> — preserve supplementary host-group access to shared datasets and workspaces. The canonical array adds every host group except the primary GID.
• HOME, USER, LOGNAME, and cache redirects — keep frameworks from writing to image-owned locations such as /root after the user override. Prepare these directories on the writable mount before launch. USER/LOGNAME are load-bearing, not cosmetic: an arbitrary --user UID has no /etc/passwd entry in the image, and torch 2.x calls getpass.getuser() at import (torch/_dynamo → inductor cache-dir setup) — with neither env var set the container crashes with KeyError: 'getpwuid(): uid not found: <uid>' before any workload code runs. Any non-empty name satisfies it; the name does not need to exist in the image.
• -v host:container — bind mount; the command references container paths only.
• -e VAR — passthrough from parent shell (no value needed if already set). Use this form for secrets.

Container name collision

docker run --name X fails if a container named X already exists. Defensive pattern before reusing a name:

docker stop my-worker 2>/dev/null; docker rm my-worker 2>/dev/null
docker run --name my-worker ...

Detached + exec pattern

For multi-step workflows on the same container (download → run → post-process), avoid restart cost:

HOST_RESULTS=/host/results
HOST_UID="$(id -u)"
HOST_GID="$(id -g)"
[ "$HOST_UID" -ne 0 ] || { echo "Refusing writable Docker launch as UID 0" >&2; exit 1; }
HOST_IDENTITY_ARGS=(--user "$HOST_UID:$HOST_GID")
for group_id in $(id -G); do
  [ "$group_id" = "$HOST_GID" ] || HOST_IDENTITY_ARGS+=(--group-add "$group_id")
done
mkdir -p "$HOST_RESULTS/.tao-runtime/home/.cache"/{huggingface,torch,triton,torchinductor,matplotlib}

docker run -d --name <worker> \
  --gpus all --shm-size=8g \
  "${HOST_IDENTITY_ARGS[@]}" \
  -v <host-data>:/data \
  -v "$HOST_RESULTS:/results" \
  -e HOME=/results/.tao-runtime/home \
  -e USER="$(id -un)" -e LOGNAME="$(id -un)" \
  -e XDG_CACHE_HOME=/results/.tao-runtime/home/.cache \
  -e HF_HOME=/results/.tao-runtime/home/.cache/huggingface \
  -e TORCH_HOME=/results/.tao-runtime/home/.cache/torch \
  -e TRITON_CACHE_DIR=/results/.tao-runtime/home/.cache/triton \
  -e TORCHINDUCTOR_CACHE_DIR=/results/.tao-runtime/home/.cache/torchinductor \
  -e MPLCONFIGDIR=/results/.tao-runtime/home/.cache/matplotlib \
  --entrypoint sh \
  <image> -c "tail -f /dev/null"

docker exec <worker> <step_1>
docker exec <worker> <step_2>

docker stop <worker> && docker rm <worker>

Pull-if-missing idiom

docker image inspect <image> >/dev/null 2>&1 || docker pull <image>

Labels for discovery

Tag containers for filtered listing later:

docker run --label tao-toolkit ...
docker ps --filter 'label=tao-toolkit'

Mount patterns

The container expects its data at conventional paths defined by the image (often /data, /results, /workspace/checkpoints). The host side is arbitrary. The command inside docker run references container paths only.

Writable-mount ownership invariant

For every writable bind mount, run as the submitting host UID:GID by default. Pre-creating the mount root is not sufficient when a root container can create deeper 0755 directories: deletion is controlled by the parent-directory permissions, so those subtrees still become inaccessible to the host user. Container --rm and docker rm remove container state only; neither deletes or repairs bind-mounted checkpoints.

An image may run as root only when its documentation or a preflight proves that host-user execution is incompatible. Treat this as an explicit launch exception. Isolate its writable outputs and, after every terminal exit or cancellation, normalize ownership before another experiment starts. For an image with /bin/sh and chown, the post-run repair is:

HOST_UID="$(id -u)"
HOST_GID="$(id -g)"
docker run --rm --user 0:0 --entrypoint /bin/sh \
  -v /host/results:/owned-output \
  <same-approved-image> \
  -c 'chown -R "$1:$2" /owned-output' sh "$HOST_UID" "$HOST_GID"

Apply the repair to every writable output/cache mount. If the agent cannot run or verify the ownership normalization, it must not use the root-required exception. Never substitute chmod 777 as the normal fix.

Env-var conventions

Common passthrough vars for TAO-style workloads (the calling skill declares which it needs):

• NGC_KEY — nvcr.io pulls; some runtimes also read at runtime
• HF_TOKEN — gated HuggingFace model downloads
• AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_ENDPOINT_URL — S3 I/O inside the container
• WANDB_API_KEY — optional W&B logging

Use -e VAR (no =value) when the var is in the parent shell. Avoid placing secrets on the command line.

Alternative GPU selection: -e NVIDIA_VISIBLE_DEVICES=0,1 (or all) and -e NVIDIA_DRIVER_CAPABILITIES=all instead of --gpus. The --gpus flag is preferred on standard x86 hosts; the env-var form is older and is what runtime=nvidia (Tegra/Jetson) requires.

Container inspection

docker ps                                # running containers only
docker ps -a                             # all containers, including exited
docker ps --filter status=running --format '{{.Names}} {{.Image}}'
docker logs <name_or_id>                 # stdout/stderr
docker logs -f <name_or_id>              # follow (tail -f equivalent)
docker logs --tail 100 <name_or_id>      # last N lines
docker inspect <name_or_id>              # full config, mounts, env, network, state (JSON)
docker inspect --format '{{.State.Status}}' <name_or_id>
docker stats                             # live CPU/mem/network/block I/O
docker stats --no-stream                 # one snapshot, non-interactive

docker inspect is the canonical source of truth for a container's mounts, env, cmd, network, and exit code. Use it to debug why a container isn't behaving as expected.

Image management

docker pull <image>
docker image ls
docker system df                # Docker-managed image/layer/volume usage

Pull once per host; docker run reuses cached image. NVIDIA images are typically 5-40GB.

Split-disk data-root relocation

Some cloud GPU providers ship with a small root volume + larger ephemeral. Docker writes to /var/lib/docker on root by default — large images fill it. Check:

df -h /         # root volume size/free
lsblk           # all block devices and mount points

If / is smaller than your total image footprint and there's a larger disk mounted elsewhere, relocate before pulling images:

sudo systemctl stop docker
sudo mkdir -p <large_volume_path>/docker
sudo rsync -aP /var/lib/docker/ <large_volume_path>/docker/
sudo mv /var/lib/docker /var/lib/docker.old

sudo tee /etc/docker/daemon.json <<'EOF'
{ "data-root": "<large_volume_path>/docker" }
EOF

sudo systemctl start docker
docker info | grep 'Docker Root Dir'
sudo rm -rf /var/lib/docker.old

Networks (multi-container patterns)

For microservice containers that talk to each other by name, create a docker network and attach containers:

docker network create tao-net
docker run --network tao-net --name api ...
docker run --network tao-net --name worker ...   # can resolve `api` by name

Most TAO training workloads don't need this — single container per job.

Common error modes

could not select device driver "" with capabilities: [[gpu]] — NVIDIA Container Toolkit missing or Docker is not configured for the NVIDIA runtime. Run tao-setup-nvidia-gpu-host with --backend docker --install after user approval (append --yes for a non-interactive agent run), then restart Docker.

unauthorized: authentication required on docker pull — NGC key invalid/missing. Re-run docker login nvcr.io.

no space left on device — first identify which filesystem and storage class is full; bind-mounted training outputs are not counted by docker system df and are not fixed by pruning Docker images:

df -h / /var/lib/docker <results_root>
docker system df
docker inspect <tao-container> --format '{{json .Mounts}}'
du -xhd1 <results_root> 2>/dev/null | sort -h
find <results_root> -maxdepth 3 -printf '%u:%g %m %s %p\n' 2>/dev/null | head

For a bind mount, clean only job directories whose record is in a terminal state (tao_job_record.py get "$JOB_ID"), via a reviewed ownership repair; never assume docker system prune touches them. For Docker's own root, relocate data-root as described above. docker system prune -a --volumes is destructive and may remove unused images and volumes belonging to other workflows, so run it only after explicit user approval and a reviewed docker system df inventory.

Bus error / DataLoader worker exited unexpectedly — /dev/shm too small. Increase shared memory with --shm-size (e.g. --shm-size=16g).

permission denied on bind-mounted paths — container UID ≠ host UID, or HOME/a framework cache still points to an image-owned directory. Use the canonical host UID:GID mapping and writable HOME/cache redirects above. For a documented root-required image, complete the mandatory post-run ownership normalization before retrying.

KeyError: 'getpwuid(): uid not found: <uid>' at import of torch/torchvision — the container runs as a --user UID with no /etc/passwd entry and no USER/LOGNAME env var, so getpass.getuser() falls through to pwd.getpwuid() at import time. -e HOME=... alone does not fix it. Keep the UID:GID mapping and launch with the canonical identity env block (-e USER=... -e LOGNAME=... + writable HOME + cache redirects). Do not work around it by running as root; that recreates the root-owned-outputs hazard.

Error: No such container: <name> after docker run -d — container crashed on startup. docker ps -a shows exited; docker logs <name> for cause. Drop --rm while debugging.

Scope boundary

This skill both runs TAO jobs on Docker (§ Execution) and documents the docker how that other skills defer to. Related:

• tao-skill-bank:tao-run-on-brev — provisions a Brev instance, then defers the container-how to these same docker verbs.
• tao-skill-bank:tao-launch-workflow — the intake/routing front door and the platform-agnostic four-verb contract this skill implements.
• tao-skill-bank:tao-data-io — the storage-tier decision + staging + the compute-frame verify gate the submit verb calls.

Model and data skills produce the spec-bundle (what); this skill runs it (how).