What problem does it solve?
This skill guides engineers through migrating CUDA, Triton, or other native operator/kernel families into the cache-dit repository while preserving semantics, minimizing native and Python closure, and preventing import-time failures when optional extensions are absent.
Core Features & Use Cases
- Survey & Scope: Defines how to identify the true call chain, required native files, and the minimal closure to migrate without dragging unnecessary upstream code.
- Design & API: Helps freeze the public surface, choose repository-aligned names, and decide what remains private versus public.
- Build, Packaging & Safety: Advises optional-extension gating, delayed extension loading, packaging updates, and explicit dependency strategies.
- Validation Strategy: Recommends layered tests from import safety to low-level correctness, module behavior, and end-to-end validation.
- Example Use Case: Port a CUDA INT4 kernel from nunchaku into cache-dit with a private CUDA helper, torch.library registrations, explicit public wrappers, and targeted tests.
Quick Start
Describe the source operator and files, the intended cache-dit public API, required backends and dtypes, and the current migration status so the skill can map the minimal viable closure and validation targets.
