Documentation

Phobos is a tile-based GPU kernel language and an LLM inference engine built on it. These pages cover both: writing kernels, and running models.

Start here

Getting startedWhat Phobos is, how to install it, and a first kernel read line by line.

Inference quickstartRun a GGUF model with phobos-cli, from a prompt or as an OpenAI-compatible server.

Writing kernels

• Language guide: the language one concept at a time, from kernels and tensors to tiles, slicing, reductions and dot.
• Execution context: every ctx field, from the grid of CTAs down to a single lane, and how per-thread values work.
• Attributes: one page for each @ attribute, such as @autotune, @launch and @tensorcore.
• Examples: five kernels from the repository, explained.

Running models

• Inference: quickstart, supported models and formats, serving, sampling, benchmarking, the kernel cache and troubleshooting.

Reference

• Language spec: the complete language, rendered from SPEC.md.
• Environment variables: every PHOBOS_* variable, rendered from ENV.md.
• Benchmarks: measured results against llama.cpp and cuBLAS, rendered from BENCHMARKS.md.
• Command-line tools: the --help of every binary.