NVIDIA CUDA

Use when work targets NVIDIA GPUs for deep learning training, inference, distributed execution, CUDA/Triton kernels, or AI infra tuning. Enforces GPU-aware c...

This package appears internally consistent and intended for local GPU tuning and code review. Before running: (1) review the scripts (they are plain Python) if you have policy concerns; (2) be aware they will query GPU drivers, environment variables, and may run GPU workloads (torch.compile, NCCL/DDP smoke tests) which can be compute- and resource-intensive; (3) run in an environment where running torch/CUDA code is safe (not on production nodes with sensitive data); and (4) the skill does not request secrets, but always inspect any third-party code before executing it with elevated privileges or on shared infrastructure.

Type: OpenClaw Skill Name: nvidia-cuda Version: 0.4.0 The bundle is a professional and comprehensive toolkit for optimizing deep learning workloads on modern NVIDIA GPUs (Hopper/Blackwell). It contains legitimate diagnostic probes (cuda_env_probe.py), static analysis tools for identifying performance anti-patterns (check_training_stack.py), and various benchmarking scripts for attention kernels and distributed training. The instructions in SKILL.md are strictly aligned with the stated purpose of performance tuning and hardware assessment, and the environment variable collection is limited to non-sensitive CUDA/NCCL configuration flags. No evidence of malicious intent, data exfiltration, or harmful prompt injection was found.