Gemma 4 26B-A4B · vLLM · NVFP4

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/s480.96
Decode tok/s421.13
Peak memory (GB)108.58 (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 32 1000 900 256
# 1000/1000 prompts (clean full run), 0 errors, 560.3s. ready after 323s. TTFT median 230.5 ms, TPOT median 72.9 ms, req thr 1.785/s.

gemma-4-26b-a4b Google Gemma NVFP4 16-40B conc-32

Notes

Decode 421.13 tok/s aggregate at concurrency 32. Autoregressive NVFP4 reference point, conc 32, 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 323 s.
  • Workload: ShareGPT V3, concurrency 32. 1000/1000, 0 errors in 560.3 s (clean full run).
  • Throughput: decode 421.13 tok/s aggregate, prefill 480.96 tok/s. TTFT median 230.5 ms, TPOT median 72.9 ms, req throughput 1.785/s.
  • Memory: 108.58 GB = vLLM --gpu-memory-utilization 0.85 reservation, not the model footprint.