# NVIDIA Dynamo [](https://opensource.org/licenses/Apache-2.0) [](https://github.com/ai-dynamo/dynamo/releases/latest) NVIDIA Dynamo is a new modular inference framework designed for serving large language models (LLMs) in multi-node distributed environments. It enables seamless scaling of inference workloads across GPU nodes and the dynamic allocation of GPU workers to address traffic bottlenecks at various stages of the model pipeline. NVIDIA Dynamo also features LLM-specific capabilities, such as disaggregated serving, which separates the context (prefill) and generation (decode) steps of inference requests onto distinct GPUs and GPU nodes to optimize performance. NVIDIA Dynamo includes four key innovations: * **Smart Router**: An LLM-aware router that directs requests across large GPU fleets to minimize costly key-value (KV) cache recomputations for repeat or overlapping requests, freeing up GPUs to respond to new incoming requests * **Low-Latency Communication Library**: An inference optimized library that supports state-of-the-art GPU-to-GPU communication and abstracts complexity of data exchange across heterogenous devices and networking protocols, accelerating data transfers * **Memory Manager**: An engine that intelligently offloads and reloads inference data (KV cache) to and from lower-cost memory and storage devices using NVIDIA NIXL without impacting user experiences > [!NOTE] > This project is currently in the alpha / experimental / > rapid-prototyping stage and we are actively looking for feedback and > collaborators. ## Quick Start TODO add quick start guide here ## Building Dynamo ### Requirements Dynamo development and examples are container based. * [Docker](https://docs.docker.com/get-started/get-docker/) * [buildx](https://github.com/docker/buildx) ### Development You can build the Dynamo container using the build scripts in `container/` (or directly with `docker build`). We provide 2 types of builds: 1. `VLLM` which includes our VLLM backend using new NIXL communication library. 2. `TENSORRTLLM` which includes our TRT-LLM backend For example, if you want to build a container for the `VLLM` backend you can run ```bash ./container/build.sh ``` Please see the instructions in the corresponding example for specific build instructions. ## Running Dynamo for Local Testing and Development You can run the Dynamo container using the run scripts in `container/` (or directly with `docker run`). The run script offers a few common workflows: 1. Running a command in a container and exiting. ```bash ./container/run.sh -- python3 -c "import dynamo.runtime; help(dynamo.runtime)" ``` ```bash ./container/run.sh -it ``` 3. Mounting the local workspace and Starting an interactive shell. ```bash ./container/run.sh -it --mount-workspace ``` The last command also passes common environment variables ( `-e HF_TOKEN` ) and mounts common directories such as `/tmp:/tmp`, `/mnt:/mnt`. Please see the instructions in the corresponding example for specific deployment instructions. ## Rust Based Runtime Dynamo has a new rust based distributed runtime with implementation under development. The rust based runtime enables serving arbitrary python code as well as native rust. Please note the APIs are subject to change. ### Hello World [Hello World](./lib/bindings/python/examples/hello_world) A basic example demonstrating the rust based runtime and python bindings. ### LLM [VLLM](./examples/python_rs/llm/vllm) An intermediate example expanding further on the concepts introduced in the Hello World example. In this example, we demonstrate [Disaggregated Serving](https://arxiv.org/abs/2401.09670) as an application of the components defined in Dynamo. # Disclaimers > [!NOTE] > This project is currently in the alpha / experimental / > rapid-prototyping stage and we will be adding new features incrementally. 1. The `TENSORRTLLM` and `VLLM` containers are WIP and not expected to work out of the box. 2. Testing has primarily been on single node systems with processes launched within a single container.