Magpie

Use Magpie for three connected jobs:

1. Benchmark an inference workload and collect throughput, latency, and traces.
2. Analyze or compare GPU kernels for correctness and performance.
3. Drive an optimization loop from a benchmark bottleneck to source, candidate kernels, and end-to-end validation.

Describe only capabilities supported by the checked-out Magpie version. Do not infer support for an unverified ROCm, GPU, framework, or experimental integration.

Choose the workflow

Use a YAML config for reproducible or multi-step work. Use inline CLI arguments for small exploratory runs.

Preflight

Installing this skill does not install the Magpie application (magpie-eval). For a preparation-only request, read the supplied fixtures and reference.md, write the requested plan or configuration, and identify commands that remain unverified. Do not install packages or execute a workload when the user forbids it; a missing Magpie installation does not prevent preparing a plan.

Before executing a Magpie workload:

1. Check magpie --help in the intended Python environment (Python 3.10+). If the command is missing, try python -m Magpie --help in that same environment; a working module entry point can be used instead of the CLI. No module named Magpie means the application is not installed in that environment. For errors inside an installed Magpie package, diagnose the reported import or dependency failure instead.

2. If Magpie is not installed, install it in the intended environment before continuing:

   python -m pip install git+https://github.com/AMD-AGI/Magpie.git
   

   For an existing Magpie source checkout, use python -m pip install -e /path/to/Magpie instead. The installed skill folder and a testcase workspace are not Magpie source checkouts. Re-run magpie --help or python -m Magpie --help to verify the installation; if it fails, report the error before attempting a workload.

3. Inspect the installed interface for the selected workflow (substitute python -m Magpie if using the module entry point):

   magpie --help
   magpie analyze --help
   magpie compare --help
   magpie benchmark --help
   magpie --gpu-info
   

4. Check required tools, model access, GPU visibility, writable output space, and container or Ray access as applicable. Installing the Python package does not install ROCm/CUDA, profilers, or model weights.

5. Read the repository compatibility matrix before making version claims. Treat ROCm or hardware not listed there as unverified until tested.

6. Record the exact config, model revision, image, environment variables, GPU allocation, and Magpie commit for benchmark comparisons.

Analyze a kernel

Prefer a config when correctness or profiler settings matter:

magpie analyze --kernel-config path/to/kernel.yaml

For a quick single-kernel run:

magpie analyze path/to/kernel.hip --type hip --testcase "./run_test.sh"

Supported public kernel types are hip, cuda, pytorch, and triton. Use --no-perf only when the user wants correctness or execution validation without profiling.

Do not equate successful execution with numerical correctness. Supply a representative testcase whenever an optimized result will be accepted or rejected.

Compare kernel variants

Compare at least two implementations and identify the baseline explicitly:

magpie compare --kernel-config path/to/compare.yaml

Keep inputs, tolerances, warmup, iteration count, GPU allocation, and profiler settings identical across candidates. Reject candidates that fail correctness before considering performance rankings.

For PyTorch without a testcase, Magpie's built-in check only verifies that each result is finite; it does not prove numerical equivalence between variants. Require a testcase for numerical validation.

Benchmark inference

Prefer a checked-in benchmark config:

magpie benchmark --benchmark-config path/to/benchmark.yaml

The stable public CLI supports vllm, sglang, and atom. It supports direct docker and local run modes; use YAML configuration and the repository's Ray examples for distributed execution. Do not advertise integrations that exist only in internal enums or partial code paths as stable.

Enable profiling deliberately: profiler runs perturb latency and should not replace a clean baseline. Compare throughput, completed requests, TTFT, TPOT, ITL, and end-to-end latency using equivalent workloads.

Post-process traces with TraceLens

Enable TraceLens in the profiled benchmark YAML; torch traces are its required input:

benchmark:
  profiler:
    torch_profiler:
      enabled: true
    tracelens:
      enabled: true
      analysis_mode: inference
      analysis_stages: all
      export_format: csv

Use analysis_mode: inference for vLLM/SGLang. It splits the rank-0 trace into prefilldecode, decode, and prefill stages when available, runs TraceLens post-processing, and writes full stage reports plus compact *_kernel_roofline_simple.csv files under the benchmark workspace's tracelens/ directory. For direct PyTorch trace reporting, use analysis_mode: pytorch.

Open the compact roofline CSVs first. Rank rows by kernel_time_ms_sum or time_pct; then use roofline_bound, arithmetic intensity, achieved TFLOP/s or TB/s, and pct_roofline_mean to form an optimization hypothesis. Confirm benchmark_report.json.tracelens_analysis has outputs and no error before treating post-processing as successful. Use analysis_mode: pytorch when the task specifically needs the legacy direct single-rank or multi-rank collective reports.

Magpie's integrated TraceLens stage produces CSV/Excel analysis artifacts, not an agent-written analysis.md. If the user requests a prioritized agentic report, pass the captured trace to the separate tracelens-analysis-orchestrator skill when installed; keep that result distinct from Magpie's benchmark report.

Analyze existing traces and find source

Run standalone gap analysis with --trace-dir directly on benchmark:

magpie benchmark \
  --trace-dir path/to/torch_trace \
  --top-k 20 \
  --find-kernel-sources \
  --kernel-source-repos path/to/repository

Do not insert a gap-analysis positional token; it is not a CLI subcommand. Inspect the generated aggregate and per-rank CSVs, and preserve source-mapping confidence rather than assuming every normalized kernel name maps uniquely.

Drive the optimization loop

1. Run an unprofiled baseline benchmark and save its config and report.
2. Repeat with torch profiling and TraceLens inference post-processing enabled.
3. Review stage-level TraceLens roofline summaries to classify dominant operations and likely compute, memory, or communication limits.
4. Run gap analysis over the representative steady-state window to rank concrete kernels.
5. Select bottlenecks by total contribution, not only single-dispatch duration.
6. Map the selected kernel to source and an executable testcase.
7. Generate isolated candidate implementations; preserve the baseline.
8. Use analyze for iteration, then compare with correctness gates to rank candidates.
9. Re-run the original unprofiled benchmark with the winning candidate and the same workload. Report both kernel-level and end-to-end changes, including regressions.

Stop before claiming success if correctness is unproven, the benchmark inputs changed, the source mapping is uncertain, or the end-to-end improvement is within run-to-run noise.

Use MCP tools when available

Prefer Magpie MCP tools for structured agent workflows such as hardware inspection, kernel discovery, config generation, analyze/compare, optimization suggestions, result lookup, report comparison, Ray job management, and benchmark batches.

Do not pass a CLI analyze_report.json wrapper directly to an MCP tool that expects one result object's performance_state and performance_result. Do not assume every CLI option exists in MCP; kernel-source enrichment is currently exposed by the CLI gap-analysis path.

Additional resources

• Full CLI reference: reference.md
• Copy-paste command examples: examples.md