Gemma 4 26B-A4B · vLLM · NVFP4 + EAGLE3 · conc 16
status: done
Configuration
| Model | google/gemma-4-26B-A4B-it |
|---|---|
| Company | |
| Family | Gemma |
| Parameters | 26B / 4B (MoE) |
| Engine | vLLM + EAGLE3 (speculative decoding) |
| Quant / precision | NVFP4 |
| Why this quant | NVIDIA NVFP4 base (modelopt) + RedHatAI's official EAGLE3 speculator (RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3) — the EAGLE3 line of the crossover figure. |
| Download | nvidia/Gemma-4-26B-A4B-NVFP4 |
| Context window | 65536 |
| Input modalities | text, image (served text-only here) |
Measured results
| Prefill tok/s | 430.9 |
|---|---|
| Decode tok/s | 376.25 |
| Peak memory (GB) | 108.85 (system MemAvailable delta (10s sampling) — vLLM static KV reservation (util 0.85) + EAGLE3 head) |
| Completed | 2026-07-03 17:54 +0800 |
Full run command
VLLM_IMAGE=vllm/vllm-openai:nightly-aarch64 scripts/bench-vllm-serving.sh nvidia/Gemma-4-26B-A4B-NVFP4 65536 16 1000 700 256 \
--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, 625.4s. ready after 314s. TTFT median 256.9 ms, TPOT median 39.7 ms, req thr 1.599/s.
# SpecDecoding: mean acceptance length 2.08 (1.86-2.21) · avg draft acceptance 36% (29-40%) · per-position ~0.58/0.32/0.18.
Notes
Decode 376.25 tok/s aggregate at concurrency 16. NVFP4 + RedHatAI EAGLE3 speculator, conc 16, 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 314 s. - Workload: ShareGPT V3, concurrency 16. 1000/1000, 0 errors in 625.4 s (clean full run).
- Throughput: decode 376.25 tok/s aggregate, prefill 430.9 tok/s. TTFT median 256.9 ms, TPOT median 39.7 ms, req throughput 1.599/s.
- Spec-decode acceptance: mean acceptance length 2.08 (1.86-2.21) · avg draft acceptance 36% (29-40%) · per-position ~0.58/0.32/0.18 (num_speculative_tokens=3).
- Memory: 108.85 GB = vLLM
--gpu-memory-utilization 0.85reservation, not the model footprint.