Gemma 4 31B · vLLM · NVFP4 + EAGLE3 · conc 8

status: done

Configuration

Modelgoogle/gemma-4-31B-it
CompanyGoogle
FamilyGemma
Parameters33B (dense)
EnginevLLM + EAGLE3 (speculative decoding)
Quant / precisionNVFP4
Why this quantNVIDIA NVFP4 base (modelopt) + RedHatAI's official EAGLE3 speculator — fast quant plus lossless speculative decoding, the combination that finally clears the 31B's time cap.
Downloadnvidia/Gemma-4-31B-IT-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s119.32
Decode tok/s91.47
Peak memory (GB)107.17 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head)
Completed2026-06-23 01:13 +08

Full run command

# vllm/vllm-openai:cu130-nightly. NVFP4 base + RedHatAI EAGLE3 speculator (draft model).
docker run -d --gpus all --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface --env HF_TOKEN=*** \
  vllm/vllm-openai:cu130-nightly nvidia/Gemma-4-31B-IT-NVFP4 \
  --host 0.0.0.0 --port 8000 --max-model-len 65536 \
  --gpu-memory-utilization 0.85 --max-num-seqs 8 \
  --speculative-config '{"model":"RedHatAI/gemma-4-31B-it-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
python3 scripts/bench-serving.py --base-url http://localhost:8000 \
  --model nvidia/Gemma-4-31B-IT-NVFP4 \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 500 --max-seconds 300 --concurrency 8 --max-tokens 256

gemma-4-31b Google Gemma NVFP4 16-40B conc-8

Notes

Conc-8 point for the Gemma 4 31B EAGLE3 sweep — acceptance steady at ~2.3, 0 errors. Closes the Gemma vLLM NVFP4+EAGLE3 set. NVIDIA NVFP4 base + RedHatAI EAGLE3 speculator on vLLM (cu130-nightly), ctx 65536, conc 8.

  • Load: ready in 392 s.
  • Workload: ShareGPT V3, concurrency 8. 129/500 completed, 0 errors before the 300 s time cap.
  • Throughput: prefill 119.32 tok/s, decode 91.47 tok/s aggregate (~11.4 tok/s/stream). TTFT median 562 ms, TPOT median 79.9 ms.
  • EAGLE3 acceptance — steady ~2.3. Run-aggregate mean acceptance length ~2.18–2.57 (centered ~2.3), avg draft acceptance ~39–52% (centered ~44%), per-position ~0.61–0.74 / 0.35–0.50 / 0.21–0.33 — essentially unchanged from conc-1 (~2.4), confirming acceptance is workload-driven. Highest vLLM-EAGLE3 acceptance of the two Gemma models (vs 26b-a4b ~2.15), matching the “bigger dense model drafts better” trend seen in the MTP runs.
  • Memory: 107.2 GB = vLLM 0.85 reservation + EAGLE3 head, not footprint.
  • Closes the Gemma vLLM NVFP4+EAGLE3 sweep (26b-a4b + 31b × conc-1/8): all clean (0 errors), EAGLE3 a solid ~2.0–2.4× draft on ShareGPT, stable across concurrency — and all on cu130-nightly, since the newer nightly-aarch64 (0.23.1) can’t load Gemma-4 NVFP4 (tie_weights regression; see the conc-1 page).