code-migration

What problem does it solve?

This Skill provides a structured guide to migrate CUDA-based PyTorch code to Ascend NPU, detailing API replacements, migration patterns, and best practices to minimize manual effort and errors.

Core Features & Use Cases

• API Replacements: Map CUDA APIs to torch_npu equivalents to ensure correct device placement and execution.
• Migration Patterns: Common code patterns for initializing models, moving tensors, and training loops.
• Best Practices & Validation: Guidance on testing correctness and performance after migration across typical PyTorch workloads.

Quick Start

Start by identifying CUDA code sections, apply the CUDA-to-NPU mapping guide, and validate results on Ascend NPU using torch_npu, iterating until parity is achieved.