Gemma 4 26B-A4B · vLLM · NVFP4 · conc 256

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. c256 point of the beyond-c32 ceiling-finding extension — base survives c256 where MTP/EAGLE3 collapse.
Downloadnvidia/Gemma-4-26B-A4B-NVFP4
Context window65536
Input modalitiestext, image (served text-only here)

Measured results

Prefill tok/s1563.93
Decode tok/s1365.89
Peak memory (GB)108.7 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85); ~10 GB free at peak, FLAT across c64/c128/c256 (no draft buffers))
Completed2026-07-04 13:01 +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 256 1000 900 256
# 1000/1000 prompts (clean full run), 0 errors, 172.3s. ready after 317s. TTFT median 768.9 ms, TPOT median 164.6 ms, req thr 5.803/s. Watchdog no trip (10 GB free).

gemma-4-26b-a4b Google Gemma NVFP4 16-40B conc-256

Notes

Decode 1365.89 tok/s aggregate at concurrency 256 — base sails clean through the point where MTP collapsed. Autoregressive NVFP4 reference, top of the planned beyond-c32 sweep (context held at 65536).

  • The headline contrast of this whole extension: at c256 the MTP line collapsed (9/1000 completed, 512 errors — draft-slot scheduler cap + memory wall) while the base line runs 1000/1000, 0 errors at 1365.89 tok/s. No draft head means no max_num_scheduled_tokens=2496 throttle and no per-seq draft buffers — so base neither thrashes the scheduler nor exhausts memory.
  • Still climbing, not saturated: decode 1365.89 vs c128 (995.1) = 1.37× per doubling — base has NOT plateaued even at c256. And memory is flat at ~10 GB free (108.70 delta, essentially unchanged from c64/c128). So base is neither compute- nor memory-bound here: its true ceiling is beyond c256 (unexplored — c512 would be the next rung; capped at c256 for this sweep). Note base at c256 (1366) essentially matches MTP at its own top c128 (1380).
  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04…. Ready after 317 s. Watchdog (floor 3.5 GB) armed, did not trip.
  • Workload: ShareGPT V3, concurrency 256. 1000/1000, 0 errors in 172.3 s (no time cap). TTFT median 768.9 ms, TPOT median 164.6 ms, req throughput 5.803/s.

</content> </invoke>