Gemma 4 31B · vLLM · FP8

status: done

Configuration

Modelgoogle/gemma-4-31B-it
CompanyGoogle
FamilyGemma
Parameters33B (dense)
EnginevLLM
Quant / precisionFP8
Why this quantRedHatAI official FP8-Dynamic (compressed-tensors) — the 1-byte point on the 31B quant ladder, to isolate quant effect from engine vs the NVFP4 run.
DownloadRedHatAI/gemma-4-31B-it-FP8-Dynamic
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s174.77
Decode tok/s147.62
Peak memory (GB)106.29 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85), see Notes)
Completed2026-06-22 10:10 +08

Full run command

# vllm/vllm-openai:cu130-nightly (ENTRYPOINT ["vllm","serve"]). RedHatAI FP8-Dynamic.
docker run -d --gpus all --ipc=host -p 8000:8000 \
  -v ~/.cache/huggingface:/root/.cache/huggingface --env HF_TOKEN=*** \
  vllm/vllm-openai:cu130-nightly RedHatAI/gemma-4-31B-it-FP8-Dynamic \
  --host 0.0.0.0 --port 8000 --max-model-len 65536 \
  --gpu-memory-utilization 0.85 --max-num-seqs 32
python3 scripts/bench-serving.py --base-url http://localhost:8000 \
  --model RedHatAI/gemma-4-31B-it-FP8-Dynamic \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256

gemma-4-31b Google Gemma FP8 16-40B conc-32

Notes

FP8 on the dense 31B — and the quant tax is much smaller here than on the MoE, because this model is compute-bound, not bandwidth-bound. Google’s Gemma-4-31B, RedHatAI FP8-Dynamic on vLLM.

  • Workload: ShareGPT V3, concurrency 32. 603/1000, 0 errorshit the 15-min cap (the dense 31B stays expensive even at FP8). Loaded in 416 s.
  • Throughput (aggregate, conc 32): prefill 174.8 tok/s, decode 147.6 tok/s. TTFT median 617 ms, TPOT median 196 ms (≈5.1 tok/s/stream), req throughput 0.64/s.
  • The dense-31B ladder, and why quant moves it less than the 26B MoE:

    Quant / engine decode completed note
    Q4_K_M / llama.cpp 78.5 322/1000 KV-cliff disaster
    FP8 / vLLM 147.6 603/1000  
    NVFP4 / vLLM 167.0 672/1000 best

    On vLLM, FP8 → NVFP4 is only +13% (148 → 167) for this dense model, vs +21% and an overall on the 26B MoE. The reason: a dense 33B fires all its parameters every token, so decode is compute-bound — halving the weight bytes (FP8→NVFP4) helps the memory side but the matmul FLOPs don’t shrink, so the gain is modest. The sparse MoE activates only ~4B/token, making it weight-bandwidth-bound, where the same quant step pays off much more. Quant format helps bandwidth-bound models far more than compute-bound ones — this pair is the clean evidence.

  • Memory: 106.3 GB is the vLLM --gpu-memory-utilization 0.85 reservation, not the footprint (FP8 weights ≈ 33 GB).
  • Text path benchmarked (mm_served: false).