FBCNN Inference removing dynamic paths in forward to make the model optimised and ready for compilation

[2292amit]

Summary

Introduces network_fbcnn_opt.py - a compilation-optimized version of FBCNN with 100% backward compatibility.

Key Features

• Compilation-ready: Removes dynamic padding and conditional logic that prevent torch.compile() optimization
• Drop-in replacement: Identical API, works with existing checkpoints and scripts
• Performance: Enables significant speedup with torch.compile() while maintaining baseline performance
• Delete intermediate tensors: Deletes temporary tensors for better memory usage

Implementation

• FBCNNCore: Pure tensor operations, no dynamic shapes or branching
• FBCNN wrapper: Handles padding/QF conversion, maintains original interface
• Compilation API: Methods like compile_core(), is_compiled(), reset_to_original_core()

Testing & Documentation

• Comprehensive unit tests covering all functionality

Enables modern PyTorch optimization without breaking existing workflows.