Proposal: a portability auditor for user-defined Keras layers and models

[jabrailkhalil]

I would like to build a developer tool that checks a user's custom layers across Keras backends and produces a small, inspectable report when behavior differs. I am interested in helping model authors assess a migration, building on Keras' existing multi-backend support.

For an MVP, I would run the same custom-layer fixtures on two backends in isolated processes, comparing shapes, dtypes, outputs, and gradients with explicit tolerances. A mismatch report would retain seeds, inputs, backend versions, and a minimal runnable example. I would distinguish unsupported behavior, expected numerical differences, and likely regressions.

Would this be more useful as an ecosystem tool or a Keras-maintained utility? I can start with a small layer set, but I would like feedback on scope and tolerance policy before proposing any new core API.

I previously proposed Qwen Code's Android work, adjusted the design with its maintainers, and contributed the merged initial shell/PWA. I would use the same small-first approach here.