TAO AutoML

Run automated hyperparameter optimization for a TAO model by combining:

1. The selected model skill under skills/models/<network>/.
2. The selected platform skill under skills/platform/<platform>/.
3. AutoMLRunner, which generates recommendations, launches train jobs,

extracts metrics, and feeds results back to the optimizer.

Do not launch until model metadata, platform preflight, data visibility, credentials, image choice, and compute shape are all proven.

Reference Map

• references/skill_info.yaml: this workflow's structured metadata.
• Split detailed references: automl-preflight-concepts.md for prerequisites

and support checks; automl-intent-algorithms.md for search policy; automl-runner-configuration.md for runner/API/WandB details; automl-advanced-monitoring.md for hooks, resume, and pitfalls; and automl-examples.md for conversation examples. detailed-guide.md is only the map.

• skills/models/<network>/SKILL.md: model-specific dataset requirements, metrics,

HPO notes, checkpoint handoff, and known failures.

• skills/models/<network>/references/skill_info.yaml: train action contract,

container image, inputs, outputs, upload exclusions, and mode.

• skills/platform/<platform>/SKILL.md: selected platform preflight, credentials,

resource shape, monitoring, and cancellation.

• skills/core/tao-launch-workflow/SKILL.md: shared intake pattern for platform,

credentials, dataset visibility, image confirmation, and user confirmation.

Preflight

1. Run the shared launch intake. If the user has not chosen a platform, ask;

Brev, SLURM, Kubernetes, and Docker are equal peers.

1. Run the selected platform skill's preflight before generating runner files.
2. Verify nvidia-tao-automl imports:

bash
python -c "import tao_automl; from tao_automl.runner import AutoMLRunner; print('OK')"

If missing, show the exact install command from versions.yaml and ask before installing:

bash
SB="${TAO_SKILL_BANK_PATH:-~/tao-skills-external}"
pip install "$($SB/scripts/resolve_versions_key.py wheels.tao_automl_<platform>)"

Valid platform wheel keys are tao_automl_brev, tao_automl_slurm, tao_automl_kubernetes, tao_automl_docker, and tao_automl_all. Use all only for development machines that need every backend. Add ,llm only when the user requests LLM-guided algorithms.

Model Support Gate

Before every run:

1. Read the model SKILL.md and references/skill_info.yaml.
2. Confirm automl_enabled: true for the model or that the model skill

explicitly routes train-stage requests to AutoML.

1. Confirm <skill_dir>/schemas/train.schema.json exists and parses. This is

the AutoML search-space gate.

1. For non-TAO-Core models such as Cosmos-RL and CLIP, also require

references/spec_template_train.yaml; otherwise the runner has no complete train defaults.

1. If any gate fails, do not improvise a search space. Report the missing

package artifact.

Inputs

Collect these before runner construction:

| Input | Requirement | |---|---| | model_skill | Resolved model skill directory under skills/models/. Accept user aliases such as network_arch only after resolving them to the packaged skill directory. | | platform | One of the supported TAO platform skills. | | train_dataset / eval_dataset | Use model-specific spec keys and dataset layout. | | results_root | Local, Lustre, or S3 path appropriate for the platform. | | gpu_count, num_nodes | Respect model and platform limits. | | container_image | Resolve through model metadata and versions.yaml; show it to the user. | | automl_algorithm | Default bayesian unless user asks for another algorithm or the model skill recommends one. | | metric, direction | Prefer the model skill's validation/task metric. | | automl_budget | Recommendation count, max epochs/rungs, concurrency, or population size as required by the algorithm. |

Never ask for secret values. Verify required env vars with [ -n "$VAR_NAME" ] && echo SET || echo UNSET.

Pre-Launch Review Gate

Before launching any recommendation jobs, show a concrete launch review and get user confirmation. This gate applies to every AutoML run for every AutoML-supported model/network; it is not Cosmos-specific and must not be scoped to a single model skill. This applies even when platform and image preflight already passed. The review must include:

• model/network, platform, image, GPU/node shape, and result/workspace root
• dataset mode and concrete spec keys, including train/eval sample counts when

they can be read cheaply

• algorithm, budget, max concurrent jobs, metric, and direction
• searchable parameters and ranges, including default values when the user did

not provide an explicit search space

• exact generated recommendation configs for the initial launch batch, produced

in a review-only step before any recommendation job is submitted

• estimated runtime per recommendation and total expected wall time, with the

assumptions used

• the automatic baseline eval job id, metric value, and result path from the

post-preflight eval job, or an explicit blocker if the model has no runnable evaluate action or validation data

• the post-AutoML final evaluation plan for the selected best checkpoint/model,

including metric, dataset, and record path

If the estimate is longer than the user's stated limit or materially longer than a normal interactive run, ask whether to reduce recommendations, epochs, dataset size, validation frequency, or search space before launch. Do not hide multi-day estimates in logs.

Automatic Baseline Eval Job

After platform, image, credential, data, and model preflight pass, run the model's evaluate action once on the selected validation/eval data before submitting any AutoML recommendation jobs. This is required AutoML setup, not an optional "pretrained eval" question for the user. Use the same base model or checkpoint that the AutoML training run starts from, the model skill's evaluate spec/template, and the selected platform's normal job submission path. If the model skill recommends a smaller shape for evaluation than training, use that shape and call it out in the launch review.

Share the eval metric number with the user in the launch review before asking for confirmation to launch recommendations. If the model has no packaged evaluate action, the eval dataset is missing, or the eval job fails, stop and report the blocker instead of silently falling back to a training-loss-only AutoML run. Continue without this baseline only when the user explicitly accepts that the run will optimize a proxy metric and will not have an impact baseline.

The AutoML runner owns final evaluation of the selected best checkpoint/model. When a runnable evaluate action and validation/eval data exist, pass a final_eval_fn(best_rec, train_job_id) callback to AutoMLRunner.run. The callback must evaluate the selected best checkpoint/model with the same metric, dataset, and direction used for the baseline, store a structured record under the workspace, and return the measured metric or a dict containing metric_value and metadata such as record_path and job_id. Do not run final evaluation as an agent-side step after runner.run; the returned result should contain result["final_evaluation"] with a concrete status and reason.

Dependency And Data Preflight

If the selected workflow needs object storage or a platform CLI and the tool is missing, report the missing dependency and offer the exact install command before continuing. After user approval, rerun scripts/check_tao_launch_preflight.py with --install-missing-tools so it installs the smallest needed package and immediately retries path verification. For S3 paths, verify both credentials and path readability from the launch platform before creating runner artifacts. Do not wait for the first training container to discover a missing AWS CLI, S3 client, or unreadable URI.

For models that read large media archives or directories during every training trial, stage or extract the dataset once to storage visible from the execution platform, then point all recommendation specs at that staged path. Record the source URI, staged path, byte/file-count evidence when available, and timestamp in <workspace>/evaluations/data_staging.json. If staging is not possible, include the repeated S3 I/O risk in the pre-launch review and ask before spending a long AutoML budget on it.

When the model skill defines sample-count-sensitive constraints, enforce them before launch. Reject or cap every batch-size recommendation that would create zero training steps for the selected dataset and GPU shard count. Use scripts/check_tao_launch_preflight.py --effective-batch-limit train_annotation=<batch_size>,<shard_count> for each generated recommendation before submitting it. If a recommendation later fails because the data is too small for the effective batch size, classify it as an invalid configuration, replace or adjust it only when remaining budget exists, and report the correction in the final summary. When train sample count is known from an annotation file or cheap manifest read, pass it as automl_settings["train_sample_count"] to AutoMLRunner.run so the runner can cap impossible recommendations before submitting a job and record the adjustment in result["history"][i]["adjustments"].

Algorithm Policy

| Algorithm | Good fit | Required knobs | |---|---|---| | bayesian | Default for small/medium budgets and few parameters. | num_recommendations, metric, direction | | hyperband, asha | Many configs with cheap early rungs; ASHA supports parallelism. | max_epochs, reduction_factor, optional max_concurrent | | bohb, dehb | Mixed Bayesian/evolutionary search with multi-fidelity budgets. | same rung budget fields as Hyperband | | pbt | Long training where schedules should mutate during training. | population and generation budget | | llm, hybrid, autoresearch | User explicitly wants LLM-guided search and has an endpoint configured. | LLM endpoint config plus budget |

Prefer the model skill's recommendation over generic defaults. Avoid ASHA or Hyperband when the model skill says startup, validation, or checkpoint cost dominates short trials.

Spec And Search Space

Build specs as nested dictionaries. If a model skill lists paths in dotted notation for readability, walk the path and assign the nested leaf; do not store flat dotted strings as spec keys.

Use the packaged train schema for:

• automl_default_parameters
• automl_disabled_parameters
• valid min/max ranges
• enums, option weights, conditions, dependencies, and popular parameters

User-provided search spaces must stay inside schema constraints. For integer knobs with discrete choices, include the schema's required integer option shape instead of a loose list if the model skill calls that out.

Data source overrides are mandatory unless the model skill says the launcher can derive them. Preserve exact user-provided spec keys when the dataset uses direct annotation/media paths.

Metric Policy

Training loss is cheap but can be misleading. Prefer the model skill's task metric. Use one of these:

• Log metric: metric=<name>, direction=maximize|minimize.
• metric_extractor(logs, metric_name): parse the model's logs when the

default resolver is ambiguous.

• eval_fn(rec, train_job_id): run the model's evaluate action after each

recommendation when the user wants a downstream task metric.

Do not map kpi to a metric unless the model skill explicitly defines that mapping.

For every AutoML run with a runnable evaluate action and validation/eval data, run the automatic baseline eval job after preflight and before recommendations. The final report must compare that baseline metric, each recommendation's metric, and the selected best metric so users can see the impact of tuning. For model skills that require an eval_fn to compute the real task metric, use that evaluator instead of optimizing a convenient training loss unless the user explicitly accepts the proxy metric.

Runner Construction

Use the selected platform SDK only after its preflight passes. Construct SDKs without embedding credentials in code.

python
from pathlib import Path
from tao_automl.runner import AutoMLRunner

skill_bank = Path("<absolute-tao-skill-bank>")
model_skill = "<resolved-model-skill-directory>"
skill_dir = skill_bank / "skills" / "models" / model_skill

runner = AutoMLRunner(
    skill_dir=str(skill_dir),
    platform_sdk=sdk,
    workspace_dir="<automl_workspace>",
)

result = runner.run(
    automl_algorithm=algorithm,
    automl_settings=automl_settings,
    spec_overrides=spec_overrides,
    automl_hyperparameters=automl_hyperparameters,
    custom_param_ranges=custom_param_ranges,
    metric_extractor=metric_extractor,  # optional
    eval_fn=eval_fn,                    # optional
    final_eval_fn=final_eval_fn,        # optional but required when final eval is runnable
)

Only resume an existing workspace when the user explicitly asks to resume, continue, recover, or inspect an existing experiment. Treat a plain "run AutoML" request as a fresh run.

Monitoring

Use runner status output and the platform SDK's get_job_status, get_job_logs, and get_failure_analysis. For active jobs, report:

• recommendation id / trial id
• platform job id
• status
• current metric
• best metric so far
• selected hyperparameters for the current/best recommendation
• elapsed time and updated ETA when enough timing data exists

On failure, classify whether it is infrastructure, data visibility, image, credential, spec/schema, or model-code failure. Fix only the minimal cause and do not silently spend additional budget on repeated invalid recommendations. If a blocker is fixed during run setup, continue from the original task after showing the updated preflight/launch review instead of leaving the user to restate the request.

For LLM-based algorithms, inspect the brain logs before calling the run valid. Verify that LLM calls succeeded, proposals were generated, prior metrics were used to choose later parameter changes, and logs show keep/discard or equivalent algorithm decisions. If the brain falls back to random sampling, classify the LLM workflow as failed or blocked instead of treating it as a valid LLM-guided run.

Result Handoff

At completion:

1. Identify the best recommendation by the selected metric and direction.
2. Return the best train child job id and its result path.
3. Resolve the model checkpoint using the model skill's checkpoint metadata and

SDK helpers; do not guess filenames such as latest.

1. Report the exact search space, algorithm, budget, metric, and platform.
2. Report the automatic baseline eval job id/result path/metric, all

recommendation metrics, final evaluation status/result path/metric, failed recommendations and root causes, elapsed time, and final runtime notes.

1. If this feeds a workflow such as AutoML + DEFT, pass the winning spec

overrides and checkpoint through the workflow's declared handoff fields.

Common Pitfalls

• Do not expect ~/tao-core at runtime. Schemas and templates must be packaged

inside the model skill.

• Do not infer dataset URIs from previous runs.
• Do not precompute SDK-managed output paths; non-URI output values are routed

by the SDK.

• For SLURM, stage large datasets on Lustre rather than burning GPU allocation

time on large S3 downloads.

• For gated HuggingFace models, verify HF_TOKEN is set without reading it.
• If all recommendations fail, stop and summarize the shared root cause instead

of launching more trials.