Gemma 4 26B-A4B · vLLM · NVFP4 + EAGLE3 · conc 128

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. c128 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/s1384.59
Decode tok/s1210.22
Peak memory (GB)111.88 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head + draft buffers; ~6.6 GB free at peak)
Completed2026-07-04 13:20 +0800

Full run command

# Run under scripts/mem-watchdog.sh (floor 3.5 GB) — GB10 unified memory; see notes/BENCHMARKING.md.
VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 128 1000 900 256 \
  --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, 194.6s. ready after 287s. TTFT median 562.6 ms, TPOT median 94.0 ms, req thr 5.138/s. Watchdog no trip (6.6 GB free).
# SpecDecoding: mean acceptance length 2.11 · avg draft acceptance 37% · per-position ~0.59/0.34/0.19.

gemma-4-26b-a4b Google Gemma NVFP4 16-40B conc-128

Notes

Decode 1210.22 tok/s aggregate at concurrency 128 — still climbing. NVFP4 + RedHatAI EAGLE3 speculator, second point of the beyond-c32 ceiling-finding sweep (context held at 65536).

  • Compute still has headroom: decode 1210.22 tok/s vs c64 (888.98) = 1.36× per doubling — matching base’s deceleration curve. EAGLE3 tracks between base and MTP throughout (base 995 · EAGLE3 1210 · MTP 1380 at c128).
  • Memory tightening (like MTP, unlike base): peak 111.88 GB delta = ~6.6 GB free — the EAGLE3 draft head + per-seq draft buffers climb with concurrency (base sat at ~9.4 GB here). This is why c256 is the danger point for the spec lines but not for base. Watchdog (floor 3.5 GB) armed, did not trip at c128.
  • Acceptance flat ~37%: mean acceptance length 2.11, per-position 0.59/0.34/0.19 — consistent with c1–c64; the weak-but-stable EAGLE3 draft.
  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04…. Ready after 287 s.
  • Workload: ShareGPT V3, concurrency 128. 1000/1000, 0 errors in 194.6 s (no time cap). TTFT median 562.6 ms, TPOT median 94.0 ms, req throughput 5.138/s.

</content> </invoke>