Develop

This section is for embedding poot as a library and for contributing kernels and engine code. If you only need the HTTP server, start with Serving. For design rationale and IR details, see Architecture.

In this section

• Build and test: Enter the Nix dev shell, sanity-check the toolchain, run just recipes.
• Compute engine: Build graphs, fuse, run on CPU/GPU, or dispatch #[kernel]s without an LLM.
• Authoring kernels: Write #[kernel] functions and dispatch them through pootc. Compute engine covers how that fits next to the graph IR path.
• Embedding poot (LLM): ModelHandle load and Driver generate API for checkpoints.
• Run a model (library): Quick library generation examples.
• Supported models: Model families, checkpoint formats, and where to check backend coverage.

Backends

poot has three production backends and a fourth, lower-priority one (--backend wgpu|rocm|ptx|vulkan):

• wgpu (default): Vulkan/SPIR-V; any modern GPU; most complete continuous-batching library path
• rocm: AMD via raw HSA (loaded at run time, no build flag)
• ptx: NVIDIA via cudarc
• vulkan: the same SPIR-V kernels as wgpu through raw Vulkan; verified on AMD (RADV) only, NVIDIA and Intel Vulkan drivers are untested

Experimental:

• NPU: experimental library devices (poot-npu); not a --backend. See NPU.

See Backends for codegen and runtime details, and Picking a backend for serve-time tradeoffs.