Docker HubServing imagesOne container per model × GPU, benchmark built in.systalyze/systalyze-serve Hugging FaceMixed INT4 / FP8 Gemma-4GPTQ INT4 experts, FP8 dense layers; 3.1× smaller than BF16.systalyze/gemma-4-26B-A4B-it-ExpertsINT4-FP8 Hugging FaceGemma-4 DSpark drafterSpeculative drafter trained on the public recipe.systalyze/gemma4-26b-a4b-it-dspark Hugging FaceDSpark multidomain drafterLong-context, workload-oriented drafter.systalyze/gemma4-26b-a4b-it-dspark-multidomain GitHubOptimized vLLM forkKernel and runtime changes; branches serving-0.27.1 and dspark-gemma4-0.28.0.systalyze/vllm-sys GitHubBenchmarks and launch linesbench/run.sh driver, per-instance configs, workloads, Dockerfile.systalyze/vllm-benchmarks
# pick a tag from the Run it column of Table 1
IMG=systalyze/systalyze-serve:gemma-4-26b-a4b-it-g7
$ docker pull $IMG
$ docker run -d --name optimized --network host --gpus all \
    -v /data/models:/models $IMG
$ curl localhost:8000/health   # ready when this returns