nemo-relay-tune-adaptive-hints

от nvidia

Consume NeMo Relay adaptive hints, predictions, latency sensitivity, ACG diagnostics, or tool-parallelism guidance safely in application logic after the…

npx skills add https://github.com/nvidia/nemo-relay --skill nemo-relay-tune-adaptive-hints

Use Adaptive Hints And Guidance

Use This When

Use this skill when the adaptive plugin is already configured and the application wants to consume adaptive outputs such as model-request hints, parallelism guidance, or ACG diagnostics.

Do Not Use This When

Do not use this skill to design the first adaptive rollout or to configure the plugin from scratch. Use nemo-relay-tune-performance or nemo-relay-tune-adaptive-config first.

Focus Areas

  • adaptive_hints metadata injected into outgoing model requests.
  • Latency sensitivity and request-local guidance.
  • Parallel groups or tool-parallelism guidance.
  • Adaptive Cache Governor reports and diagnostics during rollout.

Embedded Hint Semantics

  • Adaptive hints are request-intercept behavior. They can inject metadata into a configured header or body path; the default body path is nvext.agent_hints.
  • Hint configuration includes priority, break_chain, inject_header, and inject_body_path. Lower priority values run earlier.
  • Tool parallelism can run in observe_only, inject_hints, or schedule. Use observe_only until tool idempotency and race behavior are understood.
  • ACG output is provider-specific prompt-cache planning guidance. Use more samples, raise stability thresholds, or switch to passthrough when cache planning is unstable.
  • set_latency_sensitivity(...) is a request-local execution hint, not persistent adaptive configuration.
  • Normal adaptive runtime behavior should come from explicit config objects, not environment variables. NEMO_RELAY_ACG_DEBUG is for cache-governor diagnostics and NEMO_RELAY_RUN_REDIS_TESTS is for Redis-backed tests.

Default Path

  1. Confirm adaptive telemetry and config validation are working.
  2. Identify where the hint is injected or surfaced in the chosen binding.
  3. Keep hints advisory unless the consuming API defines stronger semantics.
  4. Test behavior when no prediction or hint is available.
  5. Compare application output against the baseline after enabling hints.
  6. Escalate from hints to scheduling only after idempotency and race behavior are understood.

Failure Modes To Avoid

  • Do not make application correctness depend on a prediction being present.
  • Do not hide behavior changes behind adaptive hints without runtime evidence.
  • Do not enable scheduling for non-idempotent or order-sensitive tools.
  • Do not adjust intercept priority without checking other request intercepts.

Related Skills

  • nemo-relay-tune-performance
  • nemo-relay-tune-adaptive-config
  • nemo-relay-debug-runtime-integration

Больше skills от nvidia

compileiq-debug
nvidia
Use when something is wrong: Search() hangs, all evaluations return INVALID_SCORE, scores aren't improving, every config returns the same number, ptxas errors…
official
create-github-pr
nvidia
Create GitHub pull requests using the gh CLI. Use when the user wants to create a new PR, submit code for review, or open a pull request. Trigger keywords -…
official
diagnose-perf
nvidia
First-responder performance triage for Isaac Sim and Isaac Lab. Identifies bottleneck category (GPU-bound, CPU-bound, VRAM, loading) using nvidia-smi and…
official
eagle3-review-logs
nvidia
Review EAGLE3 pipeline experiment logs from the launcher's experiments/ directory. Summarizes pass/fail status for all 4 tasks, diagnoses failures with root…
official
nemoclaw-maintainer-cross-issue-sweep
nvidia
Сканирует другие открытые задачи, чтобы найти те, которые данный PR может исправить или случайно сломать. Выводит возможности смежных исправлений и риски противоречий с указанием файла:строки…
official
karpathy-guidelines
nvidia
Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes,…
official
fhir-basics
nvidia
Обучает агентов работе с API FHIR R4, доступным ресурсам, запросам с параметрами поиска и корректному разбору всех форматов ответов…
official
underdeclared-agent
nvidia
A helpful assistant agent
official