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

status: done

Configuration

Modelgoogle/gemma-4-26B-A4B-it
CompanyGoogle
FamilyGemma
Parameters26B / 4B (MoE)
EnginevLLM
Quant / precisionNVFP4
Why this quantNVIDIA's own NVFP4 build (modelopt) — Blackwell-native 4-bit. Autoregressive reference line (no drafter) for 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/s1142.2
Decode tok/s995.1
Peak memory (GB)109.32 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85); ~9.4 GB free at peak (no draft buffers, so roomier than the MTP/EAGLE3 lines))
Completed2026-07-04 12:52 +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
# 1000/1000 prompts (clean full run), 0 errors, 235.9s. ready after 320s. TTFT median 396.8 ms, TPOT median 117.5 ms, req thr 4.238/s. Watchdog no trip (9.4 GB free).

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

Notes

Decode 995.1 tok/s aggregate at concurrency 128 — still climbing, and with memory to spare. Autoregressive NVFP4 reference, second point of the beyond-c32 ceiling-finding sweep (context held at 65536).

  • Compute still has clear headroom: decode 995.1 tok/s vs c64 (680.93) = 1.46× per doubling — barely decelerating from c32→c64 (1.62×). Base is nowhere near a plateau at c128.
  • Roomier than the spec lines — the key contrast: peak 109.32 GB delta = ~9.4 GB free, essentially flat vs c64 (110.16). Because the base line carries no MTP/EAGLE3 draft head and no per-seq draft-token buffers, its memory doesn’t climb with concurrency the way MTP does (mtp c128 sat at only ~4.5 GB free). This is why base — alone among the gemma lines — has the headroom to attempt c256.
  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04…. Ready after 320 s. Watchdog (floor 3.5 GB) armed, did not trip.
  • Workload: ShareGPT V3, concurrency 128. 1000/1000, 0 errors in 235.9 s (no time cap). TTFT median 396.8 ms, TPOT median 117.5 ms, req throughput 4.238/s.

</content> </invoke>