GB10 TP1 (Qwen3.8 Flash Next): alpha package at `--max-concurrency 16` gives +47% KV and faster C1/C4; one-box DFlash2+A16 export does not beat MTP

Aperta
  • Qualcosa non va?
#263 1 commento 0 reazioni 0 assegnatari Vedi su GitHub

I maintainer di solito rispondono entro 1 giorno

Nessuno ha ancora preso questa issue.

Valutazione

Difficoltà
3/5
Tempo stimato
1-2 giorni
Idoneità per principianti
48/100
Tipo di issue
Documentazione
Chiarezza
Abbastanza chiara
Stato di attività
Attiva
Stack tecnologico
github, rust
Ambito
infrastructure, performance

Direzione di ricerca

Start by reading the GB10 TP1 measurements and the alpha package README mentioned in the issue. Check how the README documents --max-concurrency and --mtp-draft-tokens, then decide whether to add the c16 recommendation and its benchmark context. Done when the guidance accurately reflects the reported results and distinguishes the one-box DFlash2 comparison.

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

Descrizione

Measurements from one DGX Spark (GB10, driver 615.71.09, kernel 7.0.0-1019-nvidia), in case they help with the GB10 TP1 defaults. Each package was benched on the same box, back to back, with no other traffic. Memory settings were identical: --memory-limit=-14GiB --embedding-table-storage disk --embedding-table-cache-size 2GiB, with page cache dropped before every launch.

Bench: llm_decode_bench.py, 8k context, 30 s per concurrency, temperature 1, 512 max tokens. Coding peak is the median of 10 runs. Single trials; run-to-run noise on this box is about ±5%.

Package Launch KV tokens C1 C4 C16 Coding peak (median / max)
local-inference-lab/Qwen3.8-Flash-Next-FLARE@gb10 (alpha, c32, MTP) default 466,944 41.7 90.9 178.0 66.9 / 82.5
same --max-concurrency 16 688,128 44.6 103.3 170.9 64.2 / 78.2
own TP1 export, DFlash2 + A16, c16 --mtp-draft-tokens 7 --mtp-adaptive-verification --mtp-adaptive-verification-alpha 0.5 618,496 41.3 85.5 169.2 71.9 / 82.0

1. Lowering concurrency at launch on the alpha package. --max-concurrency 16 on the c32 alpha package raised the KV pool 47%. It was also faster at C1 (+7%) and C4 (+14%), and within noise at C16 and coding peak. Quality on it was unchanged: Hotel Lights x8 at xhigh 8/8 (temperature 0.7, top-p 0.95, seed 1000), tool calls 5/5, vision 8/8. For a single 128 GB GB10, c16 looks like the better default, or at least worth a line in the package README next to "--max-concurrency and --mtp-draft-tokens can be reduced at launch".

2. DFlash2 + A16 on one box. We exported the same recipe we use on TP2 (two Sparks, one RoCE cable):

  • base checkpoint nvidia/Qwen3.8-Flash-Next-FP4-Local;
  • export flags --speculative-method dflash2 --mtp-draft-tokens 7 --online-mxfp8 --draft-head-quantization nvfp4 --force-moe-decode-a16 --max-concurrency 16;
  • main 972e28e4 plus host-system-reserve and dflash2-tp-embedding-contiguous.

On TP2, switching to this recipe from MTP5 took the coding-peak median from 81 to 142 tok/s. On TP1 it gained 12% coding peak over the alpha at c16, lost 17% at C4, and had 10% less KV. Fixed allocation was 76.78 GiB versus the alpha's 75.72 GiB, with 109 graphs. Our reading is that most of DFlash2's TP2 win comes from fewer cross-host verification rounds, so on one box MTP remains the better choice.

3. One-box export timing on GB10. The export took 93 minutes in total. Kernel tuning took 42 minutes from a cold cache. "Compile model forwards" took 36 minutes for 79 forwards, against 476 s for 56 forwards on the TP2 export: with about 89 GB resident in the exporting process, the automatic --compile-workers budget allowed only 2 compiler processes. Warmup, capture and packing were uneventful.

Happy to rerun anything with different flags.

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.