Nemotron-3 Puzzle 75B-A9B · vLLM · NVFP4 · conc 8

status: done

Configuration

Modelnvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4
CompanyNVIDIA
FamilyNemotron
Parameters75.3B / 9.3B (MoE, hybrid Mamba-Transformer)
EnginevLLM
Quant / precisionNVFP4
Why this quantBlackwell-native FP4 (W4A4) — GB10 gets native FP4 compute; FP8 targets Hopper, BF16 (~150 GB) won't fit 121 GB. NVIDIA's own quant, ~44.5 GB. Same Nemotron-3-Super family/quant as the 120B sibling (served on cu130-nightly; older images reject its quant_algo whitelist).
Downloadnvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4
Context window65536
Input modalitiestext

Measured results

Prefill tok/s112.82
Decode tok/s94.04
Peak memory (GB)108.86 (system MemAvailable delta (10s sampling) — vLLM 0.85 reservation; real NVFP4 weights ~53.5 GB)
Completed2026-07-11 16:54 +08

Full run command

# vllm/vllm-openai:nightly-aarch64 — loads this NVFP4 checkpoint cleanly. ENTRYPOINT ["vllm","serve"].
scripts/bench-vllm-serving.sh nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 \
  65536 8 1000 900 256 --trust-remote-code --reasoning-parser nemotron_v3

nvidia-nemotron-3-puzzle-75b-a9b NVIDIA Nemotron NVFP4 41-130B Spark recipe conc-8

Notes

Nemotron-3 Puzzle 75B-A9B (NVFP4) — conc 8. Compressed variant of Nemotron-3-Super (75.3B total / 9.3B active, hybrid Mamba-Transformer MoE: 40 Mamba + 40 MoE + 8 attention blocks). Benchmarked as the DGX-Spark-suitable variant (Blackwell-native FP4).

  • Load: 535 s on nightly-aarch64. NVFP4 weights ~53.5 GB.
  • Workload: ShareGPT V3, concurrency 8. Hit the 15-min cap at 340/1000 prompts, 0 errors (912.6 s).
  • Throughput (aggregate, conc 8): prefill 112.82 tok/s, decode 94.04 tok/s; median TTFT 21.1 s. ~4.7× the conc-1 decode (19.84 → 94.04) — near-linear batch scaling on the 9.3B active params. TTFT rises (12.8 s → 21.1 s) as 8 prompts queue their prefills; still 0 errors.