Magpie 0.3.0

Magpie 0.3.0 expands inference benchmarking with AgentX workloads, persistent
server reuse, additional Radeon runners, and new model configurations. It also
improves evaluation compatibility, TraceLens integration, and automated testing.

Release highlights

• InferenceX AgentX workloads are available through the CLI and MCP server,
  including trace-driven request lengths and dedicated benchmark examples.
• Persistent Docker benchmark servers can be reused across runs, with
  client-only examples for benchmarking an already running server.
• Radeon 8060S (gfx1151) and R9700 (gfx12) benchmark runners and examples extend
  hardware coverage. Radeon 8060S SGLang runs require a compatible custom image
  and the settings described in the
  benchmarking guide.
• New benchmark configurations cover DeepSeek V4 Pro, Qwen3-Coder,
  MiniMax-M3, and Kimi-K2.6, including MI355X examples.
• TraceLens supports extension wheels in benchmark images and automatically
  clones its source when no checkout is supplied. Atom profiling and parallel
  inference trace handling have also been improved.
• lm-eval integration improves concurrency handling, tokenized requests, and
  multi-task evaluation through the completions API.
• CPU-only unit-test CI includes a 90% package coverage gate, and benchmark
  configurations can be validated through the end-to-end benchmark workflow.

Packaging and installation

• The Python package, importable package version, and documentation version are
  now 0.3.0.
• Reproducible GitHub install examples now pin the v0.3.0 release tag.
• Optional IntelliKit dependencies are pinned to the v0.1.1 revision.
• Magpie remains beta software with a lightweight core install and optional
  integration extras.

Full changelog: v0.2.0...v0.3.0