Evaluate whether an existing hot path is a credible NVIDIA Warp candidate with reproducible evidence.
npx skills add https://github.com/NVIDIA/skills --skill warp-eval
Please help me install this Agent Skill. Skill: warp-eval Source: https://github.com/NVIDIA/skills/tree/main/skills/warp-eval Command: npx skills add https://github.com/NVIDIA/skills --skill warp-eval
Teams considering NVIDIA Warp for a performance bottleneck often lack a disciplined way to decide whether a GPU port is justified. This Skill runs a gated, evidence-driven evaluation of a narrow code seam and reports measured facts instead of adoption recommendations.
Ask your agent to evaluate whether the particle neighbor-search hot path in this repository is a credible NVIDIA Warp candidate.
High-intent search queries and answers about installing and using this skill.
Point the Skill at the repository and the suspected hot path. It screens the seam against six rejection gates, asks for explicit authorization, then profiles the real application, improves the in-project baseline, prototypes a minimum Warp kernel, and benchmarks the full boundary before reporting measured facts.
Strong candidates include irregular spatial queries, particle or geometry simulation, branch-heavy loops, many small launches, and large materialized intermediates. Dense tensor algebra already handled by PyTorch or vendor libraries, intrinsically serial recurrences, and stages too small to move the target metric are screened out.
Static screening and reporting need no GPU. Measured evaluation requires an NVIDIA CUDA GPU, the target project's dependencies, and a representative workload; without them the Skill aborts before profiling. Warp measurements are CUDA-only and synchronized, and CPU-resolved Warp runs are discarded.
It aborts when a gate fires: stated CPU-only or cross-vendor deployment requirements, an unwidenable host/device boundary, dense algebra already vendor-lowered, a mature CUDA incumbent meeting the contract, policy blocking Warp's obligations, or representative evidence that the stage cannot move the requested metric.
No. The Skill reports observed facts, measurements, hypotheses, and unknowns per seam and regime, and never recommends or ranks adoption options. It stops at the report and never modifies production code.
Measured work consumes GPU, environment, and time resources, so the Skill requires explicit approval after presenting early findings, exact scope, and resource cost. The original performance request, skill invocation, or silence does not count as authorization.
cecon123Profile Rust hot paths and validate performance improvements with flamegraph and criterion.
ShmurdocCompare web performance metrics against baselines to detect regressions.
moirai-internalIntegrate optimized CUDA and Triton kernels for diffusion model inference.
VKirillRoute Python GPU compute to CuPy or NumPy with CPU fallbacks.
