[Feature flags]Seeking Guidance on PyTorch Feature Flag Design and Best Practices

[followtheart]

I’d like to learn from PyTorch’s experience in designing feature flags to enable custom implementation. Could you help me resolve the following questions, or provide related resources for my reference:

• 1. How does PyTorch design feature flags? What principles does it follow?
• 2. How should feature flags be used in the source code to prevent interference among them? Are there specific rules, constraints, or even clever techniques?
• 3. In testing, how can one ensure that all combinations of feature flags do not introduce bugs? Does it require testing every possible combination, or are there alternative assurances?

Thx a lot