Adapt Server Config

Adapt Server Config

Create versioned AIC estimate requests without silently dropping operating points. Keep adaptation separate from estimate execution.

Choose the workflow

• InferenceX DB record plus benchmark row: use the helper with --format inferencex.
• DynamoGraphDeployment with optional perf.yaml, or a concrete dynamo-ci SGLang benchmark recipe: use the helper with --format dynamo.
• Existing canonical request: use the helper with --format request to validate it.
• Any other format: follow custom-mapping.md. Do not pass it to a known-format adapter.

Adapt a known format

Run from the aiconfigurator repository root:

uv run python .agents/skills/adapt-server-config/scripts/adapt_config.py \
  --format inferencex --config configs-record.json --benchmark benchmark-row.json \
  --overrides '{"backend_version":"0.19.0"}'

uv run python .agents/skills/adapt-server-config/scripts/adapt_config.py \
  --format dynamo --deploy deploy.yaml --perf perf.yaml \
  --overrides '{"system_name":"h200_sxm"}'

The helper prints an ordered adaptation report. Preserve every outcome, including rejected points and warnings. Treat any rejection as unresolved; do not repair it with guesses.

For benchmark cookbooks or Slurm command templates, require a rendered concrete recipe or DynamoGraphDeployment. For Helm inputs, require a rendered DynamoGraphDeployment; the matching unrendered benchmark-values.yaml can supply literal toolPipeline workload points through --perf. Do not evaluate templates or infer runtime parameters.

The helper validates adapted requests against both the Python model and packaged JSON Schema. Use --output PATH to save the report.

Handle unknown formats

Read custom-mapping.md. Inspect the source, then show the user:

1. Explicit source fields.
2. Inferred canonical fields and the evidence for each inference.
3. Missing fields and assumptions.
4. Every discovered operating point in source order.

Obtain confirmation before creating canonical request JSON. After confirmation, create aic-estimate-request/1.0.0 JSON and validate it:

uv run python .agents/skills/adapt-server-config/scripts/adapt_config.py \
  --format request --request request.json

Do not add heuristic behavior to the SDK. Do not infer missing model identity, system, workload, concurrency, or speculative-token acceptance without confirmation.

Run estimates only on explicit request

Adaptation never runs an estimate. Add --run-estimate only when the user explicitly asks to execute estimates:

uv run python .agents/skills/adapt-server-config/scripts/adapt_config.py \
  --format dynamo --deploy deploy.yaml --perf perf.yaml \
  --overrides '{"system_name":"h200_sxm"}' --run-estimate

Never execute recipe shell commands. The SDK only parses literal configuration values.