GPU Optimization for Python (NVIDIA)

What it does

Treats GPU acceleration as an evidence-driven optimization rather than an automatic rewrite.

• Library selection matrix: maps your existing stack to the right tool — CuPy (NumPy/SciPy), cuDF (pandas), cuML (scikit-learn), nx-cugraph/cuGraph (NetworkX), cuCIM (scikit-image, whole-slide imaging), cuVS (vector search / RAG), KvikIO (GPUDirect Storage, S3→GPU), Warp & Newton (simulation, robotics, autodiff), Numba-CUDA (custom kernels), RAFT (low-level primitives).
• Least-disruptive-first workflow: try accelerator modes (cudf.pandas, cuml.accel, nx-cugraph) before native APIs, and custom kernels only when profiling proves it.
• Correctness gates: compare CPU vs GPU with explicit tolerances, cover NaNs, dtypes and ordering, and report recall@k for approximate nearest-neighbor swaps.
• Honest benchmarking: CUDA events / cupyx.profiler.benchmark with warm-up, nsys timelines, ncu kernel analysis, and end-to-end latency including transfers.
• Willingness to reject the port: if GPU isn't better, it names the limiting factor (problem size, transfers, silent CPU fallback, memory pressure, launch granularity).

It also flags legacy projects — cuxfilter (final release 26.06) and cuSpatial (archived at 25.04) — so new work doesn't build on dead ends.

Who it's for

• Researchers and data scientists with large array/dataframe pipelines that run for hours
• Newcomers to RAPIDS who don't know which library fits their workload
• Teams doing graph analytics, physics simulation, vector search for RAG, or medical/pathology imaging
• Anyone who "moved to GPU" and saw no speedup and needs a real diagnosis

Examples

1. Accelerate a pandas pipeline: hand it a 50M-row aggregation script — it enables cudf.pandas, checks logs for CPU fallback, rewrites only the hot path in native cuDF, and reports before/after timings including transfer cost.
2. Swap in GPU vector search: replace sklearn NearestNeighbors in a RAG retriever with cuVS, reporting recall@k against exact search so quality loss is visible.
3. Simulation kernels: JIT-compile a Python particle-simulation loop into Warp kernels, then use an nsys timeline to remove redundant host↔device round trips.

Install the skill I found on Claude Skill Mart.
Copy the skills/optimize-for-gpu folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/optimize-for-gpu/.
When it's done, tell me in one line what this skill can do.