Gemma 4 26B-A4B · vLLM · NVFP4 + EAGLE3 · conc 4

status: done

Configuration

Modelgoogle/gemma-4-26B-A4B-it
CompanyGoogle
FamilyGemma
Parameters26B / 4B (MoE)
EnginevLLM + EAGLE3 (speculative decoding)
Quant / precisionNVFP4
Why this quantNVIDIA NVFP4 base (modelopt) + RedHatAI's official EAGLE3 speculator (RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3) — the EAGLE3 line of the crossover figure.
Downloadnvidia/Gemma-4-26B-A4B-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s191.93
Decode tok/s143.9
Peak memory (GB)109.34 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head)
Completed2026-07-03 17:54 +0800

Full run command

VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 4 500 500 256 \
  --speculative-config '{"model":"RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
# 307/500 prompts (hit the time cap), 0 errors, 505.5s. ready after 345s. TTFT median 189.4 ms, TPOT median 26.1 ms, req thr 0.607/s.
# SpecDecoding: mean acceptance length 2.11 (1.95-2.30) · avg draft acceptance 37% (32-43%) · per-position ~0.59/0.33/0.19.

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

Notes

Decode 143.9 tok/s aggregate at concurrency 4. NVFP4 + RedHatAI EAGLE3 speculator, conc 4, for the crossover figure.

  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — the SINGLE image for all 24 cells of this figure. Ready after 345 s.
  • Workload: ShareGPT V3, concurrency 4. 307/500 completed, 0 errors before the 500 s time cap.
  • Throughput: decode 143.9 tok/s aggregate, prefill 191.93 tok/s. TTFT median 189.4 ms, TPOT median 26.1 ms, req throughput 0.607/s.
  • Spec-decode acceptance: mean acceptance length 2.11 (1.95-2.30) · avg draft acceptance 37% (32-43%) · per-position ~0.59/0.33/0.19 (num_speculative_tokens=3).
  • Memory: 109.34 GB = vLLM --gpu-memory-utilization 0.85 reservation, not the model footprint.