Gemma 4 12B · llama.cpp · Q8_0

status: done

Configuration

Modelgoogle/gemma-4-12B-it
CompanyGoogle
FamilyGemma
Parameters12B (dense)
Enginellama.cpp
Quant / precisionQ8_0
Why this quantggml-org Q8_0 — the top (near-lossless) rung of the 12B llama.cpp quant ladder. The 12B is gemma4_unified (vLLM/SGLang-blocked), so llama.cpp multi-quant is its only coverage axis.
Downloadggml-org/gemma-4-12B-it-GGUF
Context window65536
Input modalitiestext (served text-only here)

Measured results

Prefill tok/s111.22
Decode tok/s153.02
Peak memory (GB)47.87 (system MemAvailable delta (10s sampling))
Completed2026-06-22 11:18 +08

Full run command

# ghcr.io/ggml-org/llama.cpp:full-cuda (dispatcher → --server). GGUF under /home/gauravmm/models.
docker run --gpus all -p 8081:8081 -v /home/gauravmm/models:/models:ro \
  ghcr.io/ggml-org/llama.cpp:full-cuda \
  --server -m /models/gemma-4-12B-it-Q8_0.gguf -ngl 99 -c 65536 \
  --parallel 32 -cb --host 0.0.0.0 --port 8081
python3 scripts/bench-serving.py --base-url http://localhost:8081 \
  --model gemma-4-12B-it-Q8_0.gguf \
  --dataset benchmark_data/ShareGPT_V3_unfiltered_cleaned_split.json \
  --num-prompts 1000 --max-seconds 900 --concurrency 32 --max-tokens 256

gemma-4-12b Google Gemma Q8_0 5-15B conc-32

Notes

The near-lossless top of the 12B llama.cpp quant ladder — and, predictably, the slowest of the plain quants. Google’s Gemma-4-12B, ggml-org Q8_0.

  • Workload: ShareGPT V3, concurrency 32. 602/1000, 13 errors, hit the 15-min cap.
  • Throughput (aggregate, conc 32): prefill 111.2 tok/s, decode 153.0 tok/s. TTFT median 3.6 s, TPOT median 171 ms.
  • Heaviest weights, slowest decode — the bandwidth-bound rule, top of the ladder: Q8_0 (8.5 bits/wt) decodes 153 vs Q6_K’s 159 and Q4_K_M’s 195, and carries the largest footprint at 47.9 GB (weights ≈ 12 GB + Gemma’s global-attention KV). It’s the quality-preservation choice, not the throughput choice; on this model the Q4→Q8 step costs ~21% decode and ~7 GB for marginal quality.
  • Context: part of the 12B’s llama.cpp-only matrix (Q4_K_M / Q6_K / Q8_0 / +Google-MTP) — the 12B is gemma4_unified and unservable on the stock vLLM/SGLang images.