Gemma 4 26B-A4B · vLLM · NVFP4 + EAGLE3 · conc 16

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/s430.9
Decode tok/s376.25
Peak memory (GB)108.85 (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 16 1000 700 256 \
  --speculative-config '{"model":"RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
# 1000/1000 prompts (clean full run), 0 errors, 625.4s. ready after 314s. TTFT median 256.9 ms, TPOT median 39.7 ms, req thr 1.599/s.
# SpecDecoding: mean acceptance length 2.08 (1.86-2.21) · avg draft acceptance 36% (29-40%) · per-position ~0.58/0.32/0.18.

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

Notes

Decode 376.25 tok/s aggregate at concurrency 16. NVFP4 + RedHatAI EAGLE3 speculator, conc 16, 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 314 s.
  • Workload: ShareGPT V3, concurrency 16. 1000/1000, 0 errors in 625.4 s (clean full run).
  • Throughput: decode 376.25 tok/s aggregate, prefill 430.9 tok/s. TTFT median 256.9 ms, TPOT median 39.7 ms, req throughput 1.599/s.
  • Spec-decode acceptance: mean acceptance length 2.08 (1.86-2.21) · avg draft acceptance 36% (29-40%) · per-position ~0.58/0.32/0.18 (num_speculative_tokens=3).
  • Memory: 108.85 GB = vLLM --gpu-memory-utilization 0.85 reservation, not the model footprint.