Skip to Content
ContributingContribution guide

Contribute to Skippy

Skippy accepts changes to inference behavior, model support, tooling, documentation, and qualification evidence. Contributions must preserve the runtime, ABI, performance, and model-behavior contracts that make a serving engine safe to release.

Use an issue to establish the problem and intended approach before investing in an implementation. Small documentation fixes still use the same issue and pull request flow so the change remains reviewable and attributable.

Open an issue and pull request

  • Start by opening a GitHub issue to discuss the proposed change. Once the approach is agreed, submit a pull request linked to that issue.
  • Submit pull requests from a fork unless you are a core contributor.
  • AI-assisted and AI-generated pull requests are welcome. Contributors remain responsible for understanding, testing, and explaining the submitted changes.
  • Before merging, changes are benchmarked against the current main branch. This protects inference performance, output correctness, KV-cache behavior, memory use, and other critical serving characteristics from regressions.
  • By submitting a contribution, you agree that it will be licensed under the Apache License 2.0.

Develop and verify the change

Read the repository CONTRIBUTING.md and the closest AGENTS.md before making changes. Build and test through Just:

just format just lint just test

Run Cargo commands serially in each checkout. For changed Rust crates, run formatting, a focused check, warning-denying all-target Clippy, and full tests, including integration and documentation tests.

The pull request should explain the user-visible behavior, relevant contracts, test coverage, and benchmark comparison. A missing required fixture or device cannot count as successful qualification.

How changes are evaluated

Review checks correctness and contract preservation first. Benchmarks compare the change with the current main branch for inference performance, output behavior, KV-cache behavior, memory use, and other affected serving paths. Hardware-specific claims require explicit qualification on the named hardware.

Repository layout

PathOwnership
crates/CLI, model tooling, runtime, serving, protocol, topology, caching, and metrics
skippy-api-rs/Separate Rust facade workspace
skippy_api/C++ facade
llama_cpp/Pinned upstream metadata and ordered patch queues
tools/xtask/Typed repository automation
docs/Operator guides, contracts, designs, and qualification records

Preserve contracts

  • Keep ABI declarations, bindings, version constants, generated docs, and certification recipes consistent.
  • Keep ordered core, model-support, and generated llama patch queues authoritative.
  • Put model-specific behavior in model builders rather than family-name branches in shared stage machinery.
  • Preserve wire and hash compatibility deliberately.
  • Default tests are offline; hardware qualification must fail or queue when a required fixture or device is absent.

Contribution reference

StepRequirement
DiscussOpen a GitHub issue
SubmitOpen a linked pull request from a fork unless you are a core contributor
VerifyRun the focused Just-driven checks required by the affected code
BenchmarkCompare affected serving behavior with current main
LicenseApache License 2.0
AI assistanceAccepted; the contributor remains responsible for the change
Last updated on