Fix inference pipeline: normalization, RAS orientation, state_dict check, debug logging

Fix TF32 API

Fix: Exact import test format, TF32 set directly (no try/except), verify allocator config includes max_split_size_mb=128

Fix: PyTorch 2.9.0 cu128, simplified setup.sh with fatal errors, TF32=tf32, cuDNN always on, TORCH_CUDA_ARCH_LIST=8.9

Fixed

Production build strategy: Docker support, pinned dependencies, OOM guardrails, optimized defaults

Optimize torch.compile: default to reduce-overhead for faster first run, add compilation time warnings

Speed optimizations: GPU aggregation, max-autotune compile, increased batch/ROI, GPU utilization monitoring, quality guardrails

Fix OOM: reduce ROI depth to 64, sw_batch_size=1, CPU aggregation, fix env var timing, update TF32 API

Performance optimizations: warm-up, channels-last, TF32, GPU monitoring, memory-mapped I/O

Major performance and stability improvements: disable torch.compile by default, GPU preprocessing, ROI auto-adjustment, timeout mechanism, validation, cuDNN benchmarking, version logging, concurrency limits

Optimize for L40S GPU (48GB VRAM): increase window sizes and batch sizes for better performance

Fix checkpoint loading order: load before torch.compile to avoid _orig_mod prefix mismatch

Add torch.compile optimization for better memory efficiency

Fix GPU OOM: Add post-load memory check, use ultra-minimal settings for low memory, and improve OOM recovery

Optimize for dedicated GPU: Add L4 GPU support with higher memory thresholds and better batch sizes

Fix naming conflict: rename model.py to model_loader.py and initialize DEVICE to prevent NoneType errors