DiffusionGemma-26B-A4B · vLLM · NVFP4 · conc 1

status: done

Configuration

Modelnvidia/diffusiongemma-26B-A4B-it
CompanyNVIDIA
FamilyGemma
Parameters25.2B / 3.8B (MoE, diffusion)
EnginevLLM
Quant / precisionNVFP4
Why this quantNVIDIA's official NVFP4 (nvidia/diffusiongemma-26B-A4B-it-NVFP4, ModelOpt) discrete-diffusion Gemma-4 MoE — decodes in parallel 256-token denoising blocks (no drafter, no verify). The diffusion line of the crossover figure — fastest at low concurrency, saturates ~199 tok/s from conc 8 onward.
Downloadnvidia/diffusiongemma-26B-A4B-it-NVFP4
Context window65536
Input modalitiestext, image, video (served text-only here)

Measured results

Prefill tok/s141.01
Decode tok/s116.04
Peak memory (GB)112.02 (system MemAvailable delta (10s sampling) — NVFP4 MoE + diffusion bidirectional-attention KV (util 0.85))
Completed2026-07-04 00:33 +0800

Full run command

VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/diffusiongemma-26B-A4B-it-NVFP4 65536 1 300 300 256 \
  --trust-remote-code --attention-backend TRITON_ATTN --reasoning-parser gemma4 --tool-call-parser gemma4 --enable-auto-tool-choice
# 153/300 prompts (hit the time cap), 0 errors, 301.1s. ready after 419s. TTFT median 1950.0 ms, TPOT median 0.0 ms, req thr 0.508/s.

diffusiongemma-26b-a4b NVIDIA Gemma NVFP4 16-40B conc-1

Notes

Decode 116.04 tok/s aggregate at concurrency 1. DiffusionGemma NVFP4 (parallel 256-token denoising blocks), conc 1, for the crossover figure.

  • Image (pinned): vllm/vllm-openai:nightly-aarch64 @ sha256:e414712fdc04… — the SINGLE image for all 24 cells of this figure. Ready after 419 s.
  • Workload: ShareGPT V3, concurrency 1. 153/300 completed, 0 errors before the 300 s time cap.
  • Throughput: decode 116.04 tok/s aggregate, prefill 141.01 tok/s. TTFT median 1950.0 ms, TPOT median 0.0 ms, req throughput 0.508/s.
  • TPOT is meaningless here — decode happens in 256-token parallel denoising blocks, so the client’s per-token latency reads near-zero. Trust the aggregate decode tok/s only.
  • Memory: 112.02 GB peak (MemAvailable delta) = ~102 GB vLLM KV reservation (util 0.85) plus diffusion’s denoising/activation buffers (~10 GB) — not just the reservation. That buffer overhead on top of a 0.85 reservation is exactly what breached the ceiling at higher concurrency.