Nemotron 3.5 Bolt on GB10: intermittent "prepared model reported device error 6" mid-stream at concurrency 1

Aperta
  • Qualcosa non va?
#52 0 commenti 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
4/5
Tempo stimato
3-5 giorni
Idoneità per principianti
48/100
Tipo di issue
Bug
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
python
Ambito
backend, performance

Direzione di ricerca

Start by tracing where streaming decode reports prepared model reported device error 6, then compare that path with the intermittent KV page reuse issue in #14. Reproduce with the supplied GB10 setup and sequential streaming requests at --max-concurrency 1, using the offered debug build or request bodies to capture more detail. Done when the cause is identified and the supplied workload no longer produces mid-stream errors.

Scritto dal modello di indicizzazione a partire dal testo della issue.

Descrizione

bug model:nemotron-3.5-bolt priority:P2 triage:investigate

Summary

Streaming chat completions intermittently fail mid-stream with prepared model reported device error 6 on a single DGX Spark (GB10) serving Nemotron 3.5 Bolt at --max-concurrency 1. The first token is streamed, then the stream ends with a server_error. 4 of the first 11 requests to this server process failed, all from one agent session.

Environment

  • NVIDIA DGX Spark, GB10 (SM121), driver 580.178.04, CUDA 13.0, kernel 7.0.0-1019-nvidia
  • Flare 47581cc (2026-10-20), release build
  • Package: nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB, exported for concurrency 4
  • Command: flare run bolt-lilicorr-c4.safetensors --gpus 0 --host 0.0.0.0 --port 8000 --served-model-name nemotron-3.5-bolt --kv-cache-limit concurrency --max-concurrency 1
  • Startup: Prepare execution. concurrency 1, Plan cache. per GPU: 27.73 GiB loaded · 69.30 GiB cache budget · reserve 4GiB · KV capped at 1 contexts

Workload

One OpenCode agent session (OpenAI-compatible /v1/chat/completions, stream: true, tools defined, chat_template_kwargs.enable_thinking: false). Requests are sequential; there is never more than one in flight. Prompt sizes at the failures were about 28K–53K tokens, mostly a reused prefix.

Server log

[Fri, 23 Oct 2026 10:16:08 GMT] request_id=chatcmpl-1307976-6 stage=stream code=server_error message="prepared model reported device error 6"
[Fri, 23 Oct 2026 10:16:46 GMT] request_id=chatcmpl-1307976-7 stage=stream code=server_error message="prepared model reported device error 6"
[Fri, 23 Oct 2026 10:16:50 GMT] request_id=chatcmpl-1307976-8 stage=stream code=server_error message="prepared model reported device error 6"
[Fri, 23 Oct 2026 10:22:57 GMT] request_id=chatcmpl-1307976-11 stage=stream code=server_error message="prepared model reported device error 6"

Client-side error

{"type":"server_error","message":"prepared model reported device error 6","param":null,"code":"server_error"}

The HTTP status is 200; the error arrives in the stream after the first content token.

Notes

  • The server stays up after each error, and the next request succeeds.
  • The same error appeared on 2026-10-21 with the same package under concurrent requests.
  • Possibly related: #14 (intermittent mid-decode errors from reused KV pages on SM120). The error code here is 6, not 1.

I can provide the full request bodies or run a debug build if that helps.

Lingua principale
Rust
Stelle
800
Fork
106
Merge medio
1g 3h
PR unite (30g)
100

Preparare l'ambiente

Come iniziare

  1. Leggi tutta la issue e poi la guida ai contributi del progetto.
  2. Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
  3. Fai un fork del repository e lavora su un branch.
  4. Apri una pull request che faccia riferimento al numero della issue.

Altre issue di NVIDIA/hyperflare-inference

Tutte le issue di NVIDIA/hyperflare-inference

Issue simili

Altre issue su Rust

Ricevi le nuove issue nella tua casella

Un breve riepilogo di issue GitHub adatte ai principianti.