# Dynamo Python Bindings Python bindings for the Dynamo runtime system, enabling distributed computing capabilities for machine learning workloads. ## 🚀 Quick Start 1. Install `uv`: https://docs.astral.sh/uv/#getting-started ``` curl -LsSf https://astral.sh/uv/install.sh | sh ``` 2. Install `protoc` protobuf compiler: https://grpc.io/docs/protoc-installation/. For example on an Ubuntu/Debian system: ``` apt install protobuf-compiler ``` 3. Setup a virtualenv ``` uv venv source .venv/bin/activate uv pip install maturin ``` 4. Build and install dynamo wheel ``` maturin develop --uv ``` 5. Experimental: To allow using mistral.rs and llama.cpp via the bindings, build with feature flags: ``` maturin develop --features mistralrs,llamacpp --release ``` `--release` is optional. It builds slower but the resulting library is significantly faster. See `examples/cli/cli.py` for usage. They will both be built for CUDA by default. If you see a runtime error `CUDA_ERROR_STUB_LIBRARY` this is because the stub `libcuda.so` is earlier on the library search path than the real libcuda. Try removing the `rpath` from the library: ``` patchelf --set-rpath '' _core.cpython-312-x86_64-linux-gnu.so ``` If you include the `llamacpp` feature flag, `libllama.so` and `libggml.so` (and family) will need to be available at runtime. ## Run Examples ### Prerequisite See [README.md](../../../docs/runtime/README.md#prerequisites). ### Hello World Example 1. Start 3 separate shells, and activate the virtual environment in each ``` source .venv/bin/activate ``` 2. In one shell (shell 1), run example server the instance-1 ``` python3 ./examples/hello_world/server.py ``` 3. (Optional) In another shell (shell 2), run example the server instance-2 ``` python3 ./examples/hello_world/server.py ``` 4. In the last shell (shell 3), run the example client: ``` python3 ./examples/hello_world/client.py ``` If you run the example client in rapid succession, and you started more than one server instance above, you should see the requests from the client being distributed across the server instances in each server's output. If only one server instance is started, you should see the requests go to that server each time. ## Performance The performance impacts of synchronizing the Python and Rust async runtimes is a critical consideration when optimizing the performance of a highly concurrent and parallel distributed system. The Python GIL is a global critical section and is ultimately the death of parallelism. To compound that, when Rust async futures become ready, accessing the GIL on those async event loop needs to be considered carefully. Under high load, accessing the GIL or performing CPU intensive tasks on on the event loop threads can starve out other async tasks for CPU resources. However, performing a `tokio::task::spawn_blocking` is not without overheads as well. If bouncing many small message back-and-forth between the Python and Rust event loops where Rust requires GIL access, this is pattern where moving the code from Python to Rust will give you significant gains.