> For clean Markdown content of this page, append .md to this URL. For the complete documentation index, see https://docs.nvidia.com/dynamo/llms.txt. For full content including API reference and SDK examples, see https://docs.nvidia.com/dynamo/llms-full.txt.

# Many-Model Serving

## Docs

- [Interpreting GPU Sharing Results](https://docs.nvidia.com/dynamo/dev/many-model-serving/interpreting-gpu-sharing-results.md): How to read the dedicated vs. shared GPU benchmark for many-model serving
- [Deploy the Baseline Experiment](https://docs.nvidia.com/dynamo/dev/many-model-serving/deploy-the-baseline-experiment.md): Run 8 Qwen3-4B models on 8 A100 GPUs, one model per GPU, as the reference for the GPU sharing experiments
- [Deploy the KAI + HAMi Experiment](https://docs.nvidia.com/dynamo/dev/many-model-serving/deploy-the-kai-ha-mi-experiment.md): Run 16 Qwen3-4B models on 8 A100 GPUs, two per GPU, with KAI-Scheduler and HAMi memory caps
- [Build the KAI-Scheduler and GPU Fractioning Forks](https://docs.nvidia.com/dynamo/dev/many-model-serving/build-the-kai-scheduler-and-gpu-fractioning-forks.md): Build, push, and install the two forks that give each model a hard compute and memory share of a GPU
- [Deploy the KAI + GPU Fractions Experiment](https://docs.nvidia.com/dynamo/dev/many-model-serving/deploy-the-kai-gpu-fractions-experiment.md): Run 16 Qwen3-4B models on 8 A100 GPUs, two per GPU, each capped at 50% of the SMs and 50% of the memory
- [Per-GPU Compute and Memory Limits with MPS](https://docs.nvidia.com/dynamo/dev/many-model-serving/per-gpu-compute-and-memory-limits-with-mps.md): Give a multi-GPU application a fixed share of every GPU, and pack smaller models into the rest