v2.1.0 - Speed up profiling 2.4x and remove the broken fx profiler (#141)

🌟 Summary

THOP v2.1.0 makes PyTorch model profiling substantially faster and simpler, delivering up to 2.4× lower profiling overhead while preserving bit-exact results. ⚡

📊 Key Changes

• Faster profiling through optimized bookkeeping

  • Replaced temporary float64 tensor buffers with lightweight Python integer attributes for operation and parameter totals.
  • Converted counting functions to return native Python numbers instead of allocating tensors.
  • Combined operation counting and parameter counting into a single forward hook.
  • Calculates each module’s parameter count once during hook registration rather than during every forward pass.

• Significant performance improvements

  • Reduced isolated profiling overhead from 13.4 ms to 5.6 ms, a 2.4× speedup.
  • Across 66 Ultralytics model configurations, get_flops became approximately 1.31× faster overall.
  • Real-world gains vary because model inference time still dominates total runtime.

• Removed the broken torch.fx profiler

  • Deleted the unused and unreliable fx_profile.py implementation.
  • This reduces maintenance burden and avoids exposing users to profiler behavior that was incomplete or inconsistent.

• Improved compatibility for custom operation rules

  • Custom rules that accumulate tensor values continue to work.
  • Profiling outputs are normalized to standard Python float values.
  • Added tests covering custom operations, built-in counting rules, and expected MAC and parameter totals. ✅

• Cleaner and more efficient counting internals

  • Simplified RNN, GRU, LSTM, pooling, normalization, activation, and upsampling counters.
  • Removed deprecated calculation helpers and unnecessary NumPy, warning, and tensor dependencies.

🎯 Purpose & Impact

• Faster model analysis: Developers can measure MACs, FLOPs, and parameter counts with less profiling overhead, especially for models containing many modules or hooks. 🚀
• More responsive tooling: Ultralytics utilities that rely on profiling, such as get_flops, should complete faster without changing reported results.
• Lower runtime overhead: Fewer hooks and fewer temporary tensor allocations reduce unnecessary PyTorch bookkeeping.
• Safer results: Bit-exact validation means the speed improvements do not alter profiling totals.
• Easier custom integrations: Third-party modules can continue using custom counting rules while receiving predictable Python numeric outputs.
• Upgrade consideration: Code that directly imports or depends on the removed thop.fx_profile module will need to migrate to the standard profile() API.

What's Changed

• Speed up profiling 2.4x and remove the broken fx profiler by @glenn-jocher in #141

Full Changelog: v2.0.22...v2.1.0