nemo-relay-plugin-adaptive-tuning

por nvidia

Usa esta habilidad cuando exista instrumentación base de NeMo Relay y el usuario quiera configurar o evaluar el comportamiento adaptativo del plugin, incluida la telemetría,…

npx skills add https://github.com/nvidia/skills --skill nemo-relay-plugin-adaptive-tuning

Tune Adaptive Plugin Behavior

Use This When

Use this skill when a user has baseline NeMo Relay instrumentation and wants to improve latency, parallelism, prompt-cache behavior, or model-request behavior from runtime signals. Keep adaptive behavior measured against a known baseline.

Do Not Use This When

Do not use this skill when the application is not instrumented yet. Start with nemo-relay-instrument-calls or nemo-relay-get-started first.

Default Guidance

  • Observe first, compare against a baseline, then enable one behavior change at a time.
  • Use the adaptive plugin component rather than inventing separate tuning logic or hand-registering adaptive behavior at every call site.
  • Start with in-memory state and telemetry-only behavior for local development.
  • Move to persistent state only when learned signals must survive restarts or be shared across workers.
  • Add active behavior only after representative runtime events show what should change.

Embedded Adaptive Model

  • Adaptive behavior is configured through the first-party plugin component with kind adaptive.
  • Adaptive requires existing NeMo Relay scopes and at least one relevant managed tool or LLM lifecycle event stream because it learns from runtime signals.
  • Main configuration areas are state, telemetry, adaptive hints, tool parallelism, Adaptive Cache Governor (ACG), and rollout policy.
  • State backends are in_memory and redis.
  • Tool-parallelism modes are observe_only, inject_hints, and schedule.
  • Adaptive Cache Governor providers are passthrough, anthropic, and openai; omit ACG until prompt-cache planning is needed.
  • Helper APIs exist in Rust nemo_relay_adaptive, Python nemo_relay.adaptive, and Node.js nemo-relay-node/adaptive. Go and raw FFI are source-first or advanced surfaces.

Default Path

Use this rollout sequence:

  1. Confirm the app emits scope events and the managed tool or LLM events needed for the behavior being evaluated. Do not require both call types when the workflow uses only one.
  2. Capture a baseline for the workflow you want to improve.
  3. Enable adaptive telemetry with in-memory state.
  4. Read references/config.md when exact plugin configuration fields are needed.
  5. Run representative traffic and inspect reports or runtime events.
  6. If configuration validation fails or expected events are absent, return the diagnostics and stop. Keep the last known working configuration active.
  7. Before enabling scheduling, verify tool idempotency and race behavior. Before enabling ACG, verify that provider request payloads are stable.
  8. Enable the smallest behavior change in config.
  9. Read references/hints.md when application logic consumes adaptive hints, tool-parallelism guidance, or ACG diagnostics.
  10. Compare results against the baseline. If latency, correctness, or failure rate regresses, restore the last known working configuration and retain the sanitized diagnostics for review.

Failure Modes To Avoid

  • Do not enable scheduling before tool idempotency and race behavior are known.
  • Do not enable prompt-cache planning before provider payloads are stable.
  • Do not treat adaptive hints as mandatory instructions unless the consuming path explicitly defines that contract.
  • Do not use environment variables as the primary adaptive configuration model.
  • Do not tune from a single run or unrepresentative traffic.
  • Do not suppress or replace original tool and model errors.
  • Do not add retries until the call owner defines their safety.
  • Revert adaptive behavior when it increases the failure rate.

Load A Reference When

  • You need the exact adaptive config shape -> references/config.md
  • You need to consume adaptive hints or scheduling guidance in app logic -> references/hints.md

Use Another Skill When

  • You need to build reusable plugin behavior instead of configuring the built-in adaptive component -> nemo-relay-plugin-build

Related Skills

  • nemo-relay-get-started
  • nemo-relay-instrument-calls
  • nemo-relay-plugin-observability
  • nemo-relay-plugin-build

Más skills de nvidia

compileiq-debug
nvidia
Úsalo cuando algo esté mal: Search() se cuelga, todas las evaluaciones devuelven INVALID_SCORE, las puntuaciones no mejoran, cada configuración devuelve el mismo número, errores de ptxas…
create-github-pr
nvidia
Crear solicitudes de extracción de GitHub usando la CLI gh. Usar cuando el usuario quiera crear un nuevo PR, enviar código para revisión o abrir una solicitud de extracción. Palabras clave de activación -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Escanea otros issues abiertos para encontrar aquellos que un PR dado también podría corregir o romper accidentalmente. Genera oportunidades de corrección adyacente y riesgos de contradicción con archivo:línea…
fhir-basics
nvidia
Enseña a los agentes cómo funcionan las APIs de FHIR R4, qué recursos están disponibles, cómo consultarlos con parámetros de búsqueda y cómo analizar correctamente todos los formatos de respuesta…
compileiq-validate-result
nvidia
Usar DESPUÉS de que una Búsqueda haya finalizado y ANTES de reclamar cualquier aceleración o enviar un ACF. Carga el CSV de dump_results, extrae los mejores K candidatos (de un solo objetivo)…
changelog-audit
nvidia
Auditar el CHANGELOG.md de Warp antes de un lanzamiento: recuperar entradas perdidas, ordenar por impacto en el usuario, refinar el lenguaje de las entradas, ajustar saltos de línea y (en modo rama de lanzamiento) incrementar comparación…
maintain-dynamic-plugins
nvidia
Mantener los cargadores de plugins dinámicos de NeMo Relay, manifiestos, SDKs nativos de Rust, protocolo de trabajador gRPC, SDK de trabajador Python, documentación, pruebas y cobertura del flujo de trabajo de lanzamiento
dgx-diagnose
nvidia
Diagnostica problemas comunes de la DGX Station GB300: fallos de CUDA, direccionamiento incorrecto de GPU, errores de contenedores vLLM/SGLang, problemas de estado MIG, errores de NVLink/Fabric Manager,…