Magpie release notes

Magpie release notes#

2026-07-22

3 min read time

Applies to Linux

This topic summarizes the features available in each Magpie release. For the hardware and software versions validated for a release, see the Compatibility matrix.

Magpie 0.2.0#

Released July 21, 2026, Magpie 0.2.0 improves TraceLens inference analysis and refreshes the documentation and compatibility guidance introduced in the initial beta.

Release highlights#

  • TraceLens inference runs now produce compact, per-stage simple roofline CSV summaries for mixed prefill/decode, prefill, and decode analysis.

  • Summary filenames include the input sequence length, output sequence length, and concurrency, making results from benchmark sweeps easier to identify.

  • TraceLens runtime patch selection is now based on the framework version installed in the runtime image and the patches available in the selected TraceLens checkout, instead of a hard-coded version list.

  • Trace processing now selects non-empty inference traces, improving report generation when profiler output contains empty trace files.

  • The compatibility matrix and the benchmarking, profiling, Ray, MCP, and troubleshooting documentation have been updated for the 0.2.0 release.

Packaging and installation#

  • The Python package, importable package version, and documentation version are now 0.2.0.

  • Reproducible GitHub install examples now pin the v0.2.0 release tag.

  • Magpie remains beta software and retains the lightweight core install and optional extras introduced in 0.1.0.

Magpie 0.1.0#

The initial public beta release of Magpie establishes a lightweight, general-purpose framework for evaluating GPU kernel correctness and performance across AMD and NVIDIA hardware.

Release highlights#

This release delivers three evaluation modes, support for AMD and NVIDIA hardware, Ray-based remote execution, TraceLens trace analysis, gap analysis, a Model Context Protocol (MCP) server, and structured JSON reporting.

Packaging and installation#

  • The package version is 0.1.0, and the Python package classifier marks Magpie as beta software.

  • Core installs remain lightweight. pip install installs only Magpie’s core Python dependencies.

  • IntelliKit components are pinned to a fixed commit and moved behind optional extras instead of being installed by default.

  • The intellikit extra installs the pinned IntelliKit integrations: metrix and accordo.

  • The all extra installs broadly supported optional integrations. It currently includes MCP support.

Documentation#

  • Installation docs now distinguish core, MCP-only, IntelliKit, and full optional installs.

  • Ray documentation now clarifies that worker requirements.txt installation does not automatically install pyproject extras.

  • Missing-tool messages for Metrix and Accordo now point to the matching Magpie extras.

Evaluation modes#

Magpie 0.1.0 ships with three evaluation modes.

  • Analyze: Single-kernel evaluation against a testcase, with optional performance profiling.

  • Compare: Multi-kernel comparison and ranking against a configurable baseline to identify the fastest correct implementation.

  • Benchmark: Framework-level benchmarking for vLLM, SGLang, and Atom, with optional torch profiler and system profiler runs.

Hardware and execution#

This release supports the following hardware and execution environments.

  • Support for AMD (HIP/ROCm) and NVIDIA (CUDA) GPUs.

  • Three execution environments: local host, sandboxed container, and remote Ray cluster.

  • Hardware-aware evaluation with optional GPU power and frequency control.

  • Automatic idle-GPU selection in Benchmark mode for both AMD and NVIDIA devices.

Kernel types#

The following kernel types are supported for compilation and evaluation.

  • HIP, CUDA, PyTorch, and Triton kernels.

Profiling and trace analysis#

The following profiling and trace analysis capabilities are included.

  • Pluggable performance profiler backends: rocprof-compute (AMD), ncu (NVIDIA), and IntelliKit Metrix.

  • Optional correctness validation using a testcase or IntelliKit Accordo.

  • TraceLens integration for performance trace analysis.

  • TraceLens inference runs emit per-stage simple roofline summary CSVs for quick kernel/op review.

  • Kernel-level gap analysis from torch profiler traces, including a standalone mode that runs on existing traces.

  • Kernel source finder to locate kernel source files and test commands for AMD kernels.

Integration#

Magpie integrates with the following external systems and workflows.

  • Model Context Protocol (MCP) server exposing Magpie capabilities to AI agents.

  • Agent skill packaging for environments without MCP.

  • Structured JSON reports for pipeline integration.

Configuration#

Magpie provides the following configuration mechanisms.

  • Framework-level configuration using config.yaml.

  • Per-evaluation kernel configuration files for analyze and compare modes.

  • Benchmark configuration files for framework benchmarks.

Contents