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

status: done

Configuration

Modelgoogle/gemma-4-26B-A4B-it
CompanyGoogle
FamilyGemma
Parameters26B / 4B (MoE)
EnginevLLM
Quant / precisionNVFP4
Why this quantNVIDIA's own NVFP4 build (TensorRT-Model-Optimizer / modelopt) — Blackwell-native 4-bit. The autoregressive reference line (no drafter) for the decode-vs-concurrency 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/s779.73
Decode tok/s680.93
Peak memory (GB)110.16 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85); ~8.5 GB free at peak)
Completed2026-07-04 12:42 +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 64 1000 900 256
# 1000/1000 prompts (clean full run), 0 errors, 345.6s. ready after 312s. TTFT median 276.4 ms, TPOT median 86.9 ms, req thr 2.893/s. Watchdog no trip (8.5 GB free).

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

Notes

Decode 680.93 tok/s aggregate at concurrency 64. Autoregressive NVFP4 reference (no drafter), first point of the beyond-c32 ceiling-finding sweep (context held at 65536; splices onto the c1–c32 curve).

  • Base scales even harder than the spec lines here: decode 680.93 tok/s vs the c32 base point (421.13) = 1.62× per doubling — steeper than MTP’s c32→c64 (1.47×), because the base line has no drafter forward passes competing for compute, so more of the added batch turns into throughput. Lots of headroom left.
  • Memory ~8.5 GB free at peak (110.16 GB delta) — the fixed 0.85 KV reservation dominates. Watchdog (floor 3.5 GB) armed throughout, did not trip.
  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — same single image as the c1–c32 cells. Ready after 312 s.
  • Workload: ShareGPT V3, concurrency 64. 1000/1000, 0 errors in 345.6 s (no time cap). TTFT median 276.4 ms, TPOT median 86.9 ms, req throughput 2.893/s (clean per-stream metrics — no reasoning parser on gemma-4).

</content> </invoke>