Nemotron-3 Puzzle 75B-A9B · vLLM · NVFP4 · conc 32

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/s186.99
Decode tok/s177.79
Peak memory (GB)109.1 (system MemAvailable delta (10s sampling) — vLLM 0.85 reservation; real NVFP4 weights ~53.5 GB)
Completed2026-07-11 17:21 +08

Full run command

# vllm/vllm-openai:nightly-aarch64 — loads this NVFP4 checkpoint cleanly. ENTRYPOINT ["vllm","serve"].
# Run under scripts/mem-watchdog.sh (6 GB floor) — conc 32 piles graph-capture/activation buffers on
# the util-0.85 reservation. It did NOT trip (survived at ~12 GB free).
scripts/bench-vllm-serving.sh nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-NVFP4 \
  65536 32 1000 900 256 --trust-remote-code --reasoning-parser nemotron_v3

nvidia-nemotron-3-puzzle-75b-a9b NVIDIA Nemotron NVFP4 41-130B Spark recipe conc-32

Notes

Nemotron-3 Puzzle 75B-A9B (NVFP4) — conc 32. 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: 565 s on nightly-aarch64. NVFP4 weights ~53.5 GB. Run under mem-watchdog.sh (6 GB floor) — did not trip (peak ~109.1 GB used, ~12 GB free), so conc 32 at ctx 65536 is safe for this model.
  • Workload: ShareGPT V3, concurrency 32. Hit the 15-min cap at 647/1000 prompts, 0 errors (920.3 s).
  • Throughput (aggregate, conc 32): prefill 186.99 tok/s, decode 177.79 tok/s; median TTFT 44.9 s.
  • Sweep scaling (conc 1 → 8 → 32): decode 19.84 → 94.04 → 177.79 tok/s, prefill 35.0 → 112.82 → 186.99. Aggregate decode keeps scaling with batch (9× from conc 1 to 32) but sub-linearly past conc 8; TTFT climbs 12.8 → 21.1 → 44.9 s as more prompts queue their prefills. 0 errors at every level.