skills.rest
NVIDIA
NVIDIAOFFICIAL
Explore all skills by this author →
WebsiteGitHub

nemo-relay-plugin-adaptive-tuning

Configure and roll out adaptive plugin behavior for instrumented NeMo Relay applications.

3.2k|370|Updated Feb 25, 2026
Search Tags:#plugin configuration#prompt caching#adaptive tuning#Performance optimization#Telemetry#nemo relay#tool parallelism
One-click install
npx skills add https://github.com/NVIDIA/skills --skill nemo-relay-plugin-adaptive-tuning
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: nemo-relay-plugin-adaptive-tuning
Source: https://github.com/NVIDIA/skills/tree/main/skills/nemo-relay-plugin-adaptive-tuning
Command: npx skills add https://github.com/NVIDIA/skills --skill nemo-relay-plugin-adaptive-tuning

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve?

Teams with existing NeMo Relay instrumentation need a measured, safe way to improve latency, tool parallelism, and prompt-cache behavior from runtime signals without destabilizing production traffic.

Core Features & Use Cases

  • Adaptive Plugin Configuration: Set up the built-in adaptive plugin component with state backends (in-memory or Redis), telemetry, adaptive hints, tool parallelism, and Adaptive Cache Governor (ACG) across Python, Node.js, and Rust bindings.
  • Measured Rollout Guidance: Follow a baseline-first sequence—capture a baseline, enable telemetry, run representative traffic, then enable one behavior change at a time with a rollback path.
  • Hint Consumption: Consume adaptive hints, tool-parallelism guidance, and ACG diagnostics safely in application logic via the nvext.agent_hints body path.
  • Use Case: A production deployment shows low prompt-cache hit rates; use this Skill to plan an ACG rollout, verify provider payload stability, compare cache hit rates against the baseline, and keep a rollback path ready.

Quick Start

Ask your agent to configure the NeMo Relay adaptive plugin with in-memory state and telemetry enabled, starting in observe-only mode against your existing instrumentation baseline.

Frequently Asked Questions about nemo-relay-plugin-adaptive-tuning

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I configure the NeMo Relay adaptive plugin?▼

Add a top-level adaptive ComponentSpec to the shared plugin config, start with an in-memory state backend, enable telemetry first, then validate the config before initializing the plugin system. Python uses nemo_relay.adaptive.AdaptiveConfig, Node.js uses nemo-relay-node/adaptive helpers, and Rust uses nemo_relay_adaptive.

How do I enable tool parallelism in NeMo Relay safely?▼

Start with tool_parallelism in observe_only mode, capture a baseline, and run representative traffic. Only escalate to inject_hints or schedule mode after verifying tool idempotency and race behavior, and revert if latency, correctness, or failure rate regresses.

When should I not use adaptive tuning in NeMo Relay?▼

Do not use adaptive tuning when the application is not instrumented yet; start with nemo-relay-get-started or nemo-relay-instrument-calls first. Adaptive behavior requires existing scopes and at least one managed tool or LLM lifecycle event stream to learn from runtime signals.

What is the Adaptive Cache Governor in NeMo Relay?▼

The Adaptive Cache Governor (ACG) is an adaptive plugin section that produces provider-specific prompt-cache planning guidance, with passthrough, anthropic, and openai providers. Enable it only after provider request payloads are stable, and use NEMO_RELAY_ACG_DEBUG for diagnostics.

How do I consume adaptive hints in my application?▼

Adaptive hints are injected into a configured header or body path, defaulting to nvext.agent_hints. Treat hints as advisory unless the consuming API defines stronger semantics, and never make application correctness depend on a prediction being present.

Should I use in-memory or Redis state for the adaptive plugin?▼

Start with the in_memory backend for local development and first rollouts. Move to Redis-backed state only when learned signals must survive restarts or be shared across workers.

Related Skills

Explore More Software Engineering Skills→
mpsuessermpsuesser
22

effect-pubsub-event-bus

Build a typed publish/subscribe event bus using Effect PubSub and Stream.

Community
Intermediate
adityasingh2400adityasingh2400
2

canary

Monitor live apps post-deploy for console errors and performance regressions.

Community
Advanced
TheBushidoCollectiveTheBushidoCollective
187

relay-fragments-patterns

Compose Relay fragments to manage component-level data dependencies in React.

Official
Advanced
ruvnetruvnet
70.1k

llm-config

Configure RuVLLM local inference with model selection, MicroLoRA fine-tuning, and SONA adaptation.

Community
Intermediate