MUSA S5000 PyTorch Development

This skill guides Python/PyTorch development on the local Moore Threads MTT S5000 platform. It is intended for CUDA-centric repositories that need one maintainable codebase across NVIDIA CUDA and Moore Threads MUSA.

Related global skills have distinct ownership:

• musa-pytorch-python: Python environments, package indexes, and wheel compatibility;
• musa-local-server: this server's installed Toolkit paths, driver, hardware, and dynamic-library diagnosis;
• musa-ubuntu2204-install: portable driver/Toolkit installation on another Ubuntu 22.04 host.

Repository instructions remain authoritative. Read the active repository's AGENTS.md, environment documentation, and experiment rules before editing or running anything. Never use this skill to bypass a repository's restrictions on GPU runs, data preparation, artifact paths, commits, or pushes.