Tao Launch Workflow

TAO Workflow Launch Intake

Use this skill before launching any TAO workflow or model action.

Quick Start

Run the platform helper, ask for platform and monitoring preferences, then run the selected platform detail helper before asking for credentials.

Non-Negotiable Launch Gate

This gate is model-agnostic. Apply it to every TAO model, data action, and application workflow before launching side-effecting work.

Do not create runner scripts, launch scripts, compatibility shims, workspace folders, state files, logs, or dependency-install side effects until the launch preflight passes.

Preflight passes only after all of these are true:

1. The execution platform is selected from the packaged platform helper.
2. Platform credentials and required credential groups are satisfied.
3. Model-specific credentials are satisfied.
4. The default container image is resolved from packaged model/action metadata, shown to the user, and either confirmed or replaced by an explicit image=<override>.
5. The platform access check succeeds from the launch host.
6. Dataset inputs are mapped to concrete spec keys and verified from the selected platform's point of view.
7. Required compute shape fields from the model/workflow skill are known.
8. Required local tools for the selected data/platform path are present, or the user approved installing the smallest missing dependency and preflight was rerun.
9. A launch review with image, platform, datasets, compute shape, expected runtime, and any generated/default configuration changes has been shown and confirmed by the user. For AutoML, the launch review must explicitly state recommendation count/budget, max concurrency, algorithm, metric, direction, and searched parameters/ranges even when defaults are used.

If any item is missing, ask for the missing input and stop before generating artifacts. This applies to AutoML, normal train/eval/infer/export/TRT, and DEFT/application workflows.

When preflight work clears a blocker, keep track of the original user request. After the fix, rerun the relevant preflight and continue toward that request; do not stop at "blocker fixed" unless the user explicitly asked only for the repair.

Initial Questions

After the user confirms what they want to do, ask for the execution platform using the packaged helper. Do not scan platform docs, skill folders, or config folders to build the choices.

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/list_tao_platforms.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} --format text

Then ask:

• Which supported platform should run this workflow?
• Should I monitor the run in this chat? Monitoring means I keep polling the backend/job logs after launch and report progress until the job finishes, fails, or you ask me to stop, even if the job stays queued for hours or days. If disabled, I launch the job, give you the job id/log path, and stop polling. Default: monitor in chat.
• How often should I post status? Default: every 5 minutes. Use 1-2 minutes for smoke tests, 5 minutes for normal training, or 10-15 minutes for long runs.

Use long_running_enabled=true and status_interval_minutes=5 when the user accepts the defaults.

When monitoring is enabled, do not send a final summary just because several polls have elapsed or the job is still PENDING. Keep the turn attached and emit status every status_interval_minutes until a terminal state or explicit user stop/detach request. If the runtime environment cannot keep the chat turn open, say that clearly and leave a durable watcher/log path; do not imply that chat updates will continue after the turn ends.

Final-answer rule: a final response ends chat-side monitoring. While long_running_enabled=true and any launched job is non-terminal, status messages must be sent as in-progress updates and the agent must continue polling. Only send a final response when the workflow reaches terminal state, the user explicitly asks to detach/stop monitoring, or the runtime genuinely cannot keep the turn open; in that last case, say it is a runtime limitation and provide the exact durable status command/log path.

Missing-Input Prompt Shape

When asking for launch inputs, include concrete examples and both dataset input modes. Do not ask only for "dataset root".

Use this structure and adapt spec keys to the selected model/action:

I need these launch inputs before I can create specs or runner files:

1. Execution platform: brev, slurm, local-docker, or kubernetes.

2. Dataset inputs. You can provide either mode:
   A) Root mode: give train/eval roots and I map required files automatically.
      Example Cosmos-RL:
      train_root=/lustre/fsw/.../cosmos/train
      -> custom.train_dataset.annotation_path=train_root/annotations.json
      -> custom.train_dataset.media_path=train_root
   B) Direct spec mode: give the exact config/spec parameters yourself.
      Example:
      custom.train_dataset.annotation_path=/lustre/fsw/.../train_annotations.json
      custom.train_dataset.media_path=/lustre/fsw/.../videos_train.tar.gz
      custom.val_dataset.annotation_path=/lustre/fsw/.../eval_annotations.json
      custom.val_dataset.media_path=/lustre/fsw/.../eval_videos/

   Platform examples:
   - SLURM/Lustre: /lustre/fsw/.../data/train or lustre:///lustre/fsw/.../data/train
   - Brev/Kubernetes: s3://bucket/path/train and s3://bucket/path/eval
   - local-docker: /data/tao/<model>/train or file:///data/tao/<model>/eval

3. Container image. I will resolve the default from packaged model metadata and
   show it before launch, for example:
   default image for <model>/<action>: <resolved container image>
   Use this image, or provide image=<override> to pin a different TAO build.

4. Compute shape required by the model, for example GPUs/nodes.

5. Required credentials from platform/model docs, for example HF_TOKEN for
   gated Hugging Face models.

6. Monitoring preference. By default I monitor in this chat and post progress
   every 5 minutes; choose 1-2 minutes for smoke tests or 10-15 minutes for
   long training.

Container Image Confirmation

Before creating specs, runner scripts, workspaces, logs, state files, or submitting a job, resolve the image for the selected model/action:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/resolve_tao_image.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} \
  --model <network> --action <action> --format text

If the helper is unavailable, read skills/models/<network>/config.json through SkillBank().get_model_config(network_arch). Resolve image fields in this order:

1. actions.<action>.container_image
2. actions.<action>.image
3. top-level container_image
4. top-level image

Show the exact image and ask:

Container image for <network>/<action>:
default=<resolved image>

Use this image, or provide image=<override>?

If the user accepts, pass the resolved image as the job image. If the user overrides, require a non-empty image reference and pass that value instead. Do not silently launch on the default image. This confirmation applies to training, AutoML recommendations, evaluation, inference, export, TensorRT engine generation, and application workflows that submit TAO containers.

Credential Filtering

After the user chooses a platform, get the credential list for only that platform:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/list_tao_platforms.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} \
  --platform <platform> --format text

Ask only for credentials returned by that command, plus model-specific credentials from the selected model skill. Do not ask for Brev credentials on SLURM, Kubernetes, or local Docker. Do not ask for SLURM credentials on Brev, Kubernetes, or local Docker. Ask S3 credentials only when the selected platform and the dataset/result URIs require s3:// access. Credentials may already be present in the process environment or in a user-approved secret env file such as ~/.tao/secrets.env or ~/.config/tao/.env; source such files only when needed and never print, grep, cat, paste, or log their contents. Verify only variable presence.

For initial launch intake, ask for required credentials and required credential groups only. Treat the helper's optional credentials/settings section as reference material; do not request those values unless their only_when condition applies, the selected workflow cannot proceed without them, or the user asks to customize that setting.

When the helper output includes a "Required credential groups" section, satisfy one credential from each group before proceeding. Explain each requested value using the helper's description and "How to get it" text.

For SLURM, user-facing prompts should ask for SSH_KEY_PATH first. Mention SSH_AUTH_SOCK only if the user says they already use an SSH agent.

Dependency Remediation

If a required CLI/library is missing, say exactly what is missing and why it is needed, then ask before installing. Examples:

• S3 dataset or results path -> require an S3-capable client such as aws.
• SDK-backed platform launch -> require the platform-specific nvidia-tao-sdk[...] extra.
• Local Docker SDK path -> require the Docker Python client and the configured Docker network.

After user approval and installation, rerun the same preflight. Do not create runner files or launch jobs between the failed check and the rerun.

Dataset Intake

Accept dataset inputs in either mode:

• Dataset root mode: the user gives train/eval/calibration roots, and the model skill maps required files by convention. Example for Cosmos-RL train: custom.train_dataset.annotation_path=<root>/annotations.json and custom.train_dataset.media_path=<root>.
• Direct spec mode: the user gives exact spec-key paths when annotations, media archives, videos, or image folders live in different places. Preserve those keys directly, for example custom.train_dataset.annotation_path=/lustre/.../train_annotations.json and custom.train_dataset.media_path=/lustre/.../videos.tar.gz.

Ask for dataset examples that match the selected platform:

• SLURM: shared cluster paths such as /lustre/fsw/portfolios/<team>/<your-dir>/data/<model>/train (where <your-dir> is your per-user directory on the cluster), or direct spec paths under /lustre/....
• Brev, Kubernetes: usually s3://bucket/path/train and s3://bucket/path/eval unless the platform profile mounts shared storage.
• Local Docker: local paths visible to the Docker host, such as /data/tao/<model>/train, or direct spec paths visible inside the planned container mount.
• Remote Docker: absolute paths visible on the remote Docker host named by DOCKER_HOST, not paths on the local agent machine.

Do not assume "dataset root" is the only acceptable input. When direct spec paths are supplied, validate the exact spec paths rather than appending default filenames.

Platform Preflight

Run the selected platform's preflight checks before any launch artifact is created.

Prefer the packaged preflight helper when the needed inputs are available:

${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/check_tao_launch_preflight.py \
  --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} \
  --platform <platform> \
  --container-image <selected-image> \
  --path train_annotation=<path> \
  --path train_media=<path>

Pass exact direct spec paths when the user supplied them. For root-mode inputs, expand model-required files first, then pass those concrete annotation/media paths to the helper.

If the helper reports a missing client tool such as aws for s3:// path verification, install the smallest needed package after user approval, then rerun the same command with --install-missing-tools and do not proceed until the rerun verifies the paths.

When the selected model skill warns that large S3 media should be staged, copy or extract the data once to platform-visible storage before creating launch artifacts, then validate those staged paths with the same preflight helper. Record the source URI and staged path in the run workspace so AutoML summaries can distinguish data staging time from training/evaluation time.

For local-docker and remote-docker, always pass the selected image with --container-image after resolving container_image from skill_info.yaml/versions.yaml. The helper verifies Docker reachability, NVIDIA Container Toolkit registration, GPU memory, selected-image architecture compatibility when known, and a GPU-visible smoke container before launch. For remote-docker, pass --docker-host or export DOCKER_HOST; the helper queries GPUs and validates bind-mounted paths through the remote daemon instead of using local host state. If the selected or smoke image is not present on the target Docker host, ask before pulling it or rerun with --pull-smoke-image after approval.

When a model skill lists annotation-level required fields, pass them with --json-required-field <path-label>=<field>[,<field>...] so schema/data content issues fail during preflight rather than inside the first training container. Do not add required annotation fields from old failure history; only enforce fields documented as required by the current model skill. For local JSON/JSONL annotation paths, the helper prints records=<N>; use the train annotation count as automl_settings["train_sample_count"] for sample-count-sensitive AutoML runs before recommendations are generated. If the model skill documents a run-local patch strategy for a missing required field, create the patched copy in the current run workspace, update the spec paths to that copy, and rerun the content check before launch. Do not ask the user to mutate source datasets unless the model skill says patching is impossible.

Do not use --skip-platform-access for a real launch. That flag is only for dry environment checks or for cases where the user has already provided explicit manual proof of platform and storage access. If the helper cannot verify remote API, CLI, cluster, or object-store access, treat preflight as failed and do not generate launch artifacts.

For SLURM:

1. Require SLURM_USER, SLURM_HOSTNAME, a partition intent, and one of SSH_KEY_PATH or SSH_AUTH_SOCK. If the user says to use the cluster default partition, pass an empty partition/omit the partition directive; do not substitute a site-specific value such as batch. Use the selected platform helper's Resource defaults for runtime values. For the packaged SLURM defaults, generate launchers with SLURM_TIME_HOURS=4 and SLURM_TIMEOUT_HOURS=3.8; never invent a 12-hour default for the 4-hour partition list. Launching the orchestrator with nohup or in the background is allowed for durability, but it does not satisfy chat monitoring by itself. After launch, keep a foreground chat-side polling loop attached until terminal state or explicit detach.
2. Split comma-separated SLURM_HOSTNAME, resolve hosts where possible, and require passwordless ssh -o BatchMode=yes to at least one host.
3. If SSH fails, do not offer several equivalent choices. Ask for SSH_KEY_PATH=/path/to/private_key and show the passwordless setup steps: create a key if needed with ssh-keygen -t ed25519 -N "" -f ~/.ssh/id_ed25519; install it with ssh-copy-id -i ~/.ssh/id_ed25519.pub <SLURM_USER>@<login-host>; trust the host with ssh-keyscan -H <login-host> >> ~/.ssh/known_hosts; set chmod 600 ~/.ssh/id_ed25519; verify with ssh -o BatchMode=yes -i ~/.ssh/id_ed25519 <SLURM_USER>@<login-host> 'hostname'; then rerun with SSH_KEY_PATH=~/.ssh/id_ed25519.
4. After SSH passes, validate dataset annotation/media paths on the remote login host with test -e or an equivalent read-only command.
5. Only then create runner scripts, specs, workspaces, or submit jobs.
6. For multi-GPU Slurm jobs, rely on the SDK Slurm backend to request --gpus-per-node=<N>. Do not generate manual --gpus=<N> sbatch snippets; that can spread GPUs across nodes and leave allocated GPUs idle.
7. For full-matrix or multi-node launches, submit one smoke job first. Launch the full matrix only after the smoke reaches training, emits the requested metric/status record, and shows expected GPU utilization.

For AutoML status, prefer structured controller/brain state and job metadata (active_jobs.json, .automl/controller/*.json, result JSON, and results_dir/train/status.json) before scanning raw logs. Parse logs only as a fallback or when the user specifically asks for log-level investigation.

For local Docker, validate Docker/GPU access and local dataset paths before writing launch artifacts. For Brev and Kubernetes, validate API or cluster access plus object-storage credentials and aws s3 ls readability for s3:// inputs before writing launch artifacts. For mounted shared-storage or PVC paths on those remote platforms, require manual proof that the path is mounted into the job environment; the helper fails closed rather than accepting unverified remote mount paths.

Runtime And Configuration Review

Before any side-effecting launch, show a concise review:

• selected platform and exact container image
• GPU ids/count and nodes, including any GPUs avoided because they are already occupied
• dataset roots or direct spec paths, with sample counts when available
• important model/workflow overrides that differ from template defaults
• estimated runtime and the assumptions behind it
• monitoring interval and whether chat-side monitoring will stay attached

For AutoML, also show the algorithm, metric/direction, recommendation budget, search parameters, ranges, and generated/default recommendation details as described in skills/applications/tao-run-automl/SKILL.md. Ask for confirmation after this review. If the user supplied a time limit, flag any plan that exceeds it and offer concrete reductions before launch.