jetson-speculative-decoding

par nvidia

Ajoutez le décodage spéculatif EAGLE-3 ou par modèle d'ébauche à un serveur Jetson vLLM lorsque le TPOT est le goulot d'étranglement.

npx skills add https://github.com/nvidia/skills --skill jetson-speculative-decoding

Jetson Speculative Decoding (vLLM)

Speculative decoding lets a small "draft" model propose tokens that the target model verifies in a single forward pass, reducing per-token latency. On Jetson, the win/loss is dominated by VRAM headroom, not by the draft quality. This skill encodes the parts an LLM won't already know.

Purpose

Tune an existing Jetson vLLM deployment for faster token generation by appending the right --speculative-config and validating whether it improves single-stream decode speed.

When to use

  • TPOT/ITL is the bottleneck (TTFT is fine, output is just slow).
  • Workload is single-stream or low-concurrency (≤2). Speculation usually loses at high concurrency.
  • Jetson family is Thor or AGX Orin. Do not suggest EAGLE-3 on Orin Nano/NX — there is rarely enough VRAM headroom to host both target and draft, and you'll OOM at startup.

When NOT to use

  • High-concurrency serving (≥8): batched decode usually beats speculation; the draft model just steals VRAM.
  • Models without a published EAGLE-3 head — do not train one ad-hoc as a "fix".
  • After applying jetson-inference-mem-tune flags that already pushed --gpu-memory-utilization near the ceiling. Free at least ~2 GB first.

Prerequisites

  • A working vLLM server recipe from jetson-llm-serve.
  • Enough memory headroom for the draft model or EAGLE-3 head in addition to the target model.
  • A benchmark baseline from jetson-llm-benchmark before enabling speculation.
  • A target model with a compatible EAGLE-3 head, or a small same-family draft model for the fallback path.

Instructions

Append --speculative-config to the vllm serve command shown in jetson-llm-serve.

EAGLE-3 (preferred when a head is published for the target model):

--speculative-config '{
  "method": "eagle3",
  "model": "<eagle3-head-repo-id>",
  "num_speculative_tokens": 5,
  "draft_tensor_parallel_size": 1
}'

Draft-model (fallback — pair a small same-family model):

--speculative-config '{
  "method": "draft_model",
  "model": "<small-draft-model-repo-id>",
  "num_speculative_tokens": 4,
  "draft_tensor_parallel_size": 1
}'

Jetson-specific tuning rules

  • num_speculative_tokens: start at 5 on Thor, 3 on AGX Orin. Higher values pay off only if the draft acceptance rate is >0.6.
  • Always pair with the same vLLM runtime path used by jetson-llm-serve: upstream vLLM 0.20+ (vllm/vllm-openai:latest) or validated native vLLM 0.20+ on Thor, upstream vLLM 0.20+ on Orin JetPack 7.2 / L4T r39+, or the NVIDIA-AI-IOT vLLM image on older Orin. Do not use an Orin NVIDIA-AI-IOT vLLM image on Thor. Older runtimes may lack EAGLE-3 or the current --speculative-config shape.
  • Drop --gpu-memory-utilization by ~0.05 vs the non-speculative baseline to give the draft model headroom.

How to verify it actually helped

  1. Run jetson-llm-benchmark (vLLM path) at --concurrency 1 before and after enabling speculation.
  2. Acceptance: target ≥30% improvement in throughput_tok_s and ≥20% drop in tpot_ms_p50 at concurrency 1.
  3. If improvement is <10%, or throughput_tok_s regresses at concurrency 8, disable speculation. The draft model is costing more than it returns.

Limitations

  • Speculative decoding improves decode-heavy workloads; it does not reduce TTFT-dominated latency.
  • High concurrency can erase the benefit because continuous batching already keeps the GPU busy.
  • Orin Nano/NX usually lack enough memory headroom for both target and draft models.
  • Acceptance rate and draft overhead are model-specific, so benchmark before and after instead of assuming a speedup.

Error handling

  • If vLLM rejects --speculative-config, verify that Thor and Orin JetPack 7.2 / L4T r39+ are using vLLM 0.20+ and that older Orin is using a JetPack-matched NVIDIA-AI-IOT vLLM image; then switch back to the non-speculative serving command if the runtime still rejects it.
  • If startup OOMs, lower --gpu-memory-utilization, use a smaller draft, or disable speculation and hand off to jetson-inference-mem-tune.
  • If benchmark throughput regresses, remove --speculative-config; a bad draft path is worse than no speculation.

Hand off to

  • jetson-llm-benchmark to quantify the change.
  • jetson-inference-mem-tune if startup OOMs after enabling speculation.

Source

vLLM speculative decoding docs and the Jetson AI Lab GenAI tutorial.

Plus de skills de nvidia

compileiq-debug
nvidia
Utilisez quand quelque chose ne va pas : Search() bloque, toutes les évaluations retournent INVALID_SCORE, les scores ne s'améliorent pas, chaque configuration retourne le même nombre, erreurs ptxas…
create-github-pr
nvidia
Créer des pull requests GitHub en utilisant l'interface en ligne de commande gh. Utiliser lorsque l'utilisateur souhaite créer une nouvelle PR, soumettre du code pour révision, ou ouvrir une pull request. Mots-clés de déclenchement -…
nemoclaw-maintainer-cross-issue-sweep
nvidia
Analyse les autres problèmes ouverts pour trouver ceux qu’une PR donnée pourrait également corriger ou casser accidentellement. Génère des opportunités de correctifs adjacents et des risques de contradiction avec fichier:ligne…
fhir-basics
nvidia
Apprend aux agents comment fonctionnent les API FHIR R4, quelles ressources sont disponibles, comment les interroger avec des paramètres de recherche, et comment analyser correctement tous les formats de réponse…
compileiq-validate-result
nvidia
Utiliser APRÈS qu'une recherche soit terminée et AVANT de réclamer un accélérateur ou d'expédier un ACF. Charge le CSV dump_results, extrait les K meilleurs candidats (mono-objectif)…
changelog-audit
nvidia
Auditer le CHANGELOG.md de Warp avant une publication : récupérer les entrées perdues, trier par impact utilisateur, affiner le langage des entrées, ajuster les retours à la ligne et (en mode branche de publication) mettre à jour la comparaison…
maintain-dynamic-plugins
nvidia
Maintenir les chargeurs de plugins dynamiques NeMo Relay, les manifestes, les SDK natifs Rust, le protocole worker gRPC, le SDK worker Python, la documentation, les tests et la couverture du workflow de publication
dgx-diagnose
nvidia
Diagnostiquer les problèmes courants du DGX Station GB300 — plantages CUDA, ciblage incorrect du GPU, bugs de conteneur vLLM/SGLang, problèmes d'état MIG, erreurs NVLink/Fabric Manager,…