Gemma 4 26B-A4B · vLLM · NVFP4 + MTP

status: done

Configuration

Modelgoogle/gemma-4-26B-A4B-it
CompanyGoogle
FamilyGemma
Parameters26B / 4B (MoE)
EnginevLLM + MTP (Google assistant drafter) (speculative decoding)
Quant / precisionNVFP4
Why this quantNVIDIA NVFP4 base (modelopt) + Google's official MTP assistant drafter (google/gemma-4-26B-A4B-it-assistant) via vLLM's native gemma-4 MTP path — the near-free-drafter line of the crossover figure.
Downloadnvidia/Gemma-4-26B-A4B-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s796.2
Decode tok/s696.98
Peak memory (GB)108.92 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + Gemma4 MTP head)
Completed2026-07-03 16:32 +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 \
  --speculative-config '{"method":"mtp","model":"google/gemma-4-26B-A4B-it-assistant","num_speculative_tokens":3}'
# 1000/1000 prompts (clean full run), 0 errors, 338.5s. ready after 288s. TTFT median 333.3 ms, TPOT median 42.1 ms, req thr 2.954/s.
# SpecDecoding: mean acceptance length 2.66 (2.26-2.86) · avg draft acceptance 56% (42-62%) · per-position ~0.74/0.54/0.38.

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

Notes

Decode 696.98 tok/s aggregate at concurrency 32. NVFP4 + Google MTP assistant drafter, conc 32 — the near-free-drafter line of the crossover figure.

  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — the SINGLE image for all 24 cells of this figure. Ready after 288 s.
  • Workload: ShareGPT V3, concurrency 32. 1000/1000, 0 errors in 338.5 s (clean full run).
  • Throughput: decode 696.98 tok/s aggregate, prefill 796.2 tok/s. TTFT median 333.3 ms, TPOT median 42.1 ms, req throughput 2.954/s.
  • Spec-decode acceptance: mean acceptance length 2.66 (2.26-2.86) · avg draft acceptance 56% (42-62%) · per-position ~0.74/0.54/0.38 (num_speculative_tokens=3).
  • Memory: 108.92 GB = vLLM --gpu-memory-utilization 0.85 reservation, not the model footprint.