KLineage

Klineage is an agentic harness for accelerator kernel generation and optimization, with reusable skills extracted from expert implementations. Klineage builds upon the ability of existing harness like Codex and Claude Code. It is designed to provide minimal abstraction for artifact organization, knowledge management and benchmarking for kernel agents across hardwares.

Documentation · Paper · API reference

Quick start

Requires Python 3.12+, uv, Codex or native Claude Code, and Linux bubblewrap with user namespaces. From the repository root, install the CLI with one backend:

uv tool install '.[cuda]' --python 3.12

Use .[ascend] or .[hygon] for those backends. Each extra installs its Python dependencies; configure the matching system driver and toolchain separately. See runtime requirements. If the commands are not on PATH, run uv tool update-shell and reopen the terminal.

Codex is the default runner. Its model and provider settings come from $CODEX_HOME/config.toml (~/.codex/config.toml when CODEX_HOME is unset). These overrides are optional:

export KLINEAGE_API_BASE_URL=https://provider.example/v1
export KLINEAGE_API_KEY_ENV=MODEL_API_KEY
export KLINEAGE_MODEL=provider-model-id

When overriding the key variable, populate MODEL_API_KEY on the host.

For an Anthropic Messages-compatible endpoint, select Claude Code:

export KLINEAGE_RUNNER=claude-code
export ANTHROPIC_BASE_URL=https://provider.example/anthropic
export ANTHROPIC_MODEL=provider-model-id
export ANTHROPIC_API_KEY=your-key

Use ANTHROPIC_AUTH_TOKEN instead for bearer authentication. KLINEAGE_MODEL, KLINEAGE_API_BASE_URL and KLINEAGE_API_KEY_ENV override the Anthropic variables; clear previous overrides when switching providers. See model access.

Generate a kernel from a problem directory containing definition.json and at least one workload in workloads.jsonl:

klineage-generation problems/gemm \
  --workdir agent-workspace/gemm --timeout 600

By default, the command detects the backend and starts local evaluation. Add --server http://127.0.0.1:9100 to use an existing evaluator. It prints the accepted Kernel as JSON; sources and per-workload results remain under versions/. See Getting started for setup and output details.