Cpu_fallback for the files

[Varun-sai-500]

No description provided.

[z-x-yang]

Thanks for the CPU-fallback refactor — good direction. I can't approve this yet because there are still blocking issues:

1. CPU path still breaks in eval.py / train.py

• The CPU branch passes rank=None into Evaluator/Trainer.
• But manager code still assumes GPU-only semantics (cfg.* + rank, torch.cuda.set_device(...), many .cuda(...) usages).
• This will fail before evaluation/training starts on CPU.

2. Single-GPU device-id regression

• Single-device launch now calls main_worker(args.gpu_id, ...) / main_worker(args.start_gpu, ...).
• Manager code also adds offsets (cfg.TEST_GPU_ID + rank, cfg.DIST_START_GPU + rank).
• For non-zero GPU ids this can double-offset and select the wrong device.

3. CUDA-only timing/memory calls are still unguarded in evaluator

• torch.cuda.Event, torch.cuda.synchronize, torch.cuda.max_memory_allocated still run in code paths that need CPU compatibility.

Please fix these before merge. Suggested direction:

• Keep rank as an integer (e.g., 0) on CPU;
• Guard all CUDA-only APIs with torch.cuda.is_available();
• Keep device selection in one place to avoid double offsets.