v2.2.0 - Count functional products and extrapolate quadratic costs from stride proxies (#180)

🌟 Summary

THOP 2.2.0 improves profile() so it can count more attention-related operations and estimate model cost from smaller inputs more reliably—reducing the need for custom counting rules or full-size profiling runs.

📊 Key Changes

• Counts functional matrix products: On PyTorch 1.13 and later, profile() now detects operations such as @, torch.matmul, torch.bmm, and—when available—scaled_dot_product_attention. Products inside a module with its own counting rule are not counted again.
• Improves stride-based estimates: Models that may have quadratic costs, such as attention-based models or models using custom operation rules, are profiled at several smaller sizes. THOP checks whether those measurements support a quadratic estimate and profiles at the target size if they do not.
• Adds min_cells: This new profile() argument lets users start proxy measurements at a larger size, which can help with models whose costs change behavior below a certain input size.
• Updates guidance and maintenance details: The English and Chinese READMEs now describe the profiling behavior, and repository instructions and CLA workflow handling were corrected. Documentation links were also cleaned up.

🎯 Purpose & Impact

• More complete operation counts: Functional attention and matrix products that module hooks could miss are now included, giving users a clearer picture of model compute.
• More useful estimates with safeguards: Smaller proxy inputs can speed up profiling, while validation and target-size fallbacks help avoid unreliable extrapolations.
• Less manual setup: Users may no longer need to write their own rules just to count common functional matrix products.
• No model architecture changes: This release focuses on THOP’s profiling behavior, documentation, and development workflows.

What's Changed

• Correct AGENTS.md Ruff target version and add intro by @glenn-jocher in #175
• Correct inaccurate AGENTS.md claims by @glenn-jocher in #176
• Remove trailing slashes and update redirected Ultralytics URLs by @pderrenger in #177
• Accept normalized CLA signature comments by @glenn-jocher in #178
• Document the three-number PyPI release policy by @glenn-jocher in #179
• Count functional products and extrapolate quadratic costs from stride proxies by @glenn-jocher in #180

Full Changelog: v2.1.6...v2.2.0