cupynumeric-migration-readiness

This skill is a pre-migration readiness assessor for porting NumPy code to cuPyNumeric, NVIDIA's distributed drop-in NumPy built on the Legate runtime. It answers a single question before engineers commit significant effort: which of the user's existing NumPy APIs will scale on multi-GPU cuPyNumeric, and which must be refactored first. It explicitly positions itself for use BEFORE migration, distinguishing distributed-scaling-friendly patterns from blockers such as unsupported APIs, scalar synchronization, host round-trips, object-heavy Python control flow, shape- or data-dependent branching, and in-place mutation hazards.

The assessment is emphatically static and read-only: it inspects source using Read, Grep, and Glob, cross-references a bundled cuPyNumeric API-support manifest, and reasons semantically about each idiom's expected multi-GPU scaling. It runs a five-step process that gathers context (source location, hot-path array sizes, target hardware, dominant compute pattern) with sensible defaults, classifies idioms against decision-framework gates (including a 65,536-element per-GPU floor and problem-size tiers), and produces a structured verdict of READY, LIGHT REFACTOR, SIGNIFICANT REFACTOR, or NOT RECOMMENDED with per-finding reasoning and refactor-recipe pointers. Rich reference material covers partitioning and load balance, the execution model, idioms that scale versus block, refactor recipes, and case studies. The skill deliberately declines out-of-scope requests like post-migration profiling or CUDA kernel authoring, and states it must not execute the user's code, modify files, or print secrets.

It targets scientific-computing and HPC developers, data scientists, and performance engineers weighing a GPU migration of NumPy workloads. Because it is a knowledge-driven, non-executing analysis that respects security boundaries (no code execution, no secret printing, read-only inspection), it is a benign advisory tool.