v2.0.21 - Remove leaked profiling buffers from all modules (#137)

🌟 Summary

🧹 THOP 2.0.21 makes model profiling side-effect-free while improving release reliability, documentation accuracy, and development workflows.

📊 Key Changes

• Fixed leaked profiling buffers by @raimbekovm:

  • profile() now removes temporary hooks and total_ops/total_params buffers from every module it touches.
  • Prevents profiling metadata from remaining in state_dict() or spreading through shared PyTorch modules.
  • FLOPs and parameter results remain unchanged.

• Strengthened PyPI publishing reliability:

  • Queues release workflows to avoid overlapping publish runs.
  • Retries failed PyPI uploads, including transient Sigstore, Rekor, and network failures.
  • Supports recovery of partially completed releases and safely reuses existing tags.
  • Allows SBOM uploads to replace incomplete files.

• Modernized CI/CD actions:

  • Updated GitHub Actions checkout, Python setup, and Slack integrations.
  • Replaced direct setup-uv usage with the shared Ultralytics action for more consistent behavior.

• Improved notifications and contributor workflows:

  • Slack release messages are shorter, clearer, and include clickable release, run, and PR links.
  • Replaced the third-party CLA action with the shared Ultralytics CLA workflow.
  • Added AGENTS.md guidance for AI coding agents, with CLAUDE.md linked to it.

• Updated documentation and code quality:

  • Corrected benchmark outputs and README links.
  • Standardized PyTorch imports and logging practices for Ruff 0.16.0 compatibility.
  • Updated the version to 2.0.21.

🎯 Purpose & Impact

• ✅ More trustworthy profiling: Users can profile a model without unintentionally modifying its saved state or affecting models created later in the same process.
• 🧠 Safer repeated profiling: This is especially important for complex models with shared activation modules or custom containers.
• 🚀 More dependable releases: Queuing and retry behavior reduce failed PyPI releases caused by temporary infrastructure or network issues.
• 🔄 Better recovery: Maintainers can resume incomplete releases without manually recreating tags or release metadata.
• 📚 Clearer documentation: Correct benchmark values and contribution links make THOP easier to understand and use.
• 🛠️ No expected API disruption: The release preserves profiling results and existing functionality while improving cleanup, automation, and maintainability.

What's Changed

• Improve README accuracy and links by @glenn-jocher in #125
• Bump actions/checkout from v6 to v7 in /.github/workflows by @UltralyticsAssistant in #126
• Simplify Slack notification messages by @glenn-jocher in #127
• Linkify PR numbers in publish Slack notifications by @glenn-jocher in #128
• Add AGENTS.md agent guidance and fix README drift by @glenn-jocher in #129
• Harden publish workflow recovery, notifications, and dispatch guard by @glenn-jocher in #130
• Replace third-party CLA action by @glenn-jocher in #131
• Bump slackapi/slack-github-action from v3.0.3 to v3.0.5 in /.github/workflows by @UltralyticsAssistant in #132
• Bump slackapi/slack-github-action from v3.0.5 to v4.0.0 in /.github/workflows by @UltralyticsAssistant in #133
• Use shared setup-uv action everywhere by @glenn-jocher in #134
• Bump actions/setup-python from v6 to v7 in /.github/workflows by @UltralyticsAssistant in #135
• Queue publish runs and retry PyPI upload failures by @glenn-jocher in #136
• Ruff 0.16.0 compatibility changes by @glenn-jocher in #138
• Ruff 0.16.0 compatibility changes by @glenn-jocher in #139
• Remove leaked profiling buffers from all modules by @raimbekovm in #137

New Contributors

• @raimbekovm made their first contribution in #137

Full Changelog: v2.0.20...v2.0.21