Gemma 4 26B-A4B · vLLM · NVFP4 + MTP · conc 64

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. c64 point of the beyond-c32 ceiling-finding extension.
Downloadnvidia/Gemma-4-26B-A4B-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s1175.76
Decode tok/s1027.53
Peak memory (GB)110.59 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + Gemma4 MTP head)
Completed2026-07-04 10:25 +0800

Full run command

VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 64 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, 229.2s. ready after 319s. TTFT median 423.7 ms, TPOT median 55.4 ms, req thr 4.363/s.
# SpecDecoding: mean acceptance length 2.72 (2.69-2.85) · avg draft acceptance 57% (56-62%) · per-position ~0.76/0.55/0.40.

gemma-4-26b-a4b Google Gemma NVFP4 16-40B conc-64

Notes

Decode 1027.53 tok/s aggregate at concurrency 64 — the MTP line is still climbing hard past c32. NVFP4 + Google MTP assistant drafter, first point of the beyond-c32 ceiling-finding sweep (context held at 65536 so it splices onto the c1–c32 curve).

  • Still rising, not saturated: decode 1027.53 tok/s vs the c32 point (696.98) = 1.47× per doubling — barely decelerating from the c16→c32 step (1.60×). At c64 this MTP line is nowhere near a compute plateau; the ceiling is further out.
  • Memory 110.59 GB against the 121 GB unified limit — the vLLM --gpu-memory-utilization 0.85 reservation (fixed regardless of --max-num-seqs), essentially unchanged from c32 (108.92). Raising the seq cap does not raise the reservation; headroom to the ceiling is ~10 GB.
  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — same single image as the c1–c32 cells. Ready after 319 s.
  • Workload: ShareGPT V3, concurrency 64. 1000/1000, 0 errors in 229.2 s (clean full run, no time cap). TTFT median 423.7 ms, TPOT median 55.4 ms, req throughput 4.363/s.
  • Spec-decode acceptance: mean acceptance length 2.72 · avg draft acceptance 57% · per-position ~0.76/0.55/0.40 (num_speculative_tokens=3) — flat vs the c1–c32 points, confirming acceptance is workload-driven not concurrency-driven.

</content> </invoke>