Gemma 4 26B-A4B · vLLM · NVFP4 · conc 16

status: done

Configuration

Modelgoogle/gemma-4-26B-A4B-it
CompanyGoogle
FamilyGemma
Parameters26B / 4B (MoE)
EnginevLLM
Quant / precisionNVFP4
Why this quantNVIDIA's own NVFP4 build (TensorRT-Model-Optimizer / modelopt) — Blackwell-native 4-bit. The autoregressive reference line (no drafter) for the decode-vs-concurrency crossover figure.
Downloadnvidia/Gemma-4-26B-A4B-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s312.65
Decode tok/s273.61
Peak memory (GB)109.3 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85))
Completed2026-07-03 15:15 +0800

Full run command

VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 16 1000 700 256
# 824/1000 prompts (hit the time cap), 0 errors, 711.5s. ready after 293s. TTFT median 186.1 ms, TPOT median 56.9 ms, req thr 1.158/s.

gemma-4-26b-a4b Google Gemma NVFP4 16-40B conc-16

Notes

Decode 273.61 tok/s aggregate at concurrency 16. Autoregressive NVFP4 reference point, conc 16, for the decode-vs-concurrency crossover figure.

  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — the SINGLE image for all 24 cells of this figure. Ready after 293 s.
  • Workload: ShareGPT V3, concurrency 16. 824/1000 completed, 0 errors before the 700 s time cap.
  • Throughput: decode 273.61 tok/s aggregate, prefill 312.65 tok/s. TTFT median 186.1 ms, TPOT median 56.9 ms, req throughput 1.158/s.
  • Memory: 109.30 GB = vLLM --gpu-memory-utilization 0.85 reservation, not the model footprint.