pymc

Imported from pymc-devs/pymc (AGENTS.md). Install upstream with npx skills add pymc-devs/pymc. Copyright stays with the author.

Contributing to PyMC

Prioritize useful improvements to supported modeling workflows. Correctness, performance, compatibility, maintenance, and reviewer attention all matter.

Is this a bug?

• Identify the intended contract and a realistic user impact. Distinguish wrong results with valid inputs from undefined behavior, unsupported features, and intentional design choices.
• Reproduce on current code through the actual PyMC workflow. Model gradients include pre-gradient rewrites; raw symbolic differentiation is not equivalent. Account for normal compilation, transformations, and backend behavior.
• Check the implementation, related work, and maintainer discussions. A stale branch or an already-fixed problem does not require another patch.
• Respect the division of responsibilities with PyTensor. An unstable-looking expression or an artificial execution path alone does not justify duplicating compiler stabilization or adding defensive code in PyMC.

Is fixing it an acceptable trade-off?

• Weigh the practical benefit against runtime, compilation, complexity, compatibility, and maintenance costs. Do not burden common valid uses with checks or branches for speculative edge cases.
• Assess numerical accuracy and performance across relevant ranges. Improving one example can degrade another or interfere with existing optimizations.
• Address shared causes at the appropriate layer. Prefer general solutions and coherent policy changes to accumulating local workarounds.
• Treat user-visible behavior changes as design decisions, even when the diff is small. Decline or defer changes whose benefits do not justify their costs.
• Verify automated findings independently and keep work bounded. Consolidate related findings instead of opening an issue or PR for each speculative failure. Maintainer attention is part of a contribution's cost.

Use focused, proportionate tests of the actual failing workflow. Explain the user impact, evidence, and trade-offs concisely. Distinguish a rejected proposal from work that needs separate investigation or review.