This commit adds configurations and setup scripts to resolve NCCL timeout
issues during DeepSpeed ZeRO-2 training on H200 GPUs. The changes include:
- Extended NCCL and DeepSpeed timeouts
- Optimized bucket sizes for gradient communication
- CPU and dataloader optimizations
- System shared memory improvements
- Enhanced debugging capabilities
The implementation provides:
1. DeepSpeed ZeRO-2 configuration (ds_config_zero2.json)
2. Environment setup script (setup_training_env.sh)
3. Accelerate configuration (accelerate_config.yaml)
These changes improve training stability on H200 GPUs with high-resolution
data and aggressive configurations.

...ations
When using FreeU with half-precision (torch.float16) models, PyTorch may emit
UserWarnings about experimental ComplexHalf support during FFT operations.
This change locally suppresses that specific warning in the fourier_filter
function to avoid flooding user logs while preserving behavior.
- Added warnings import
- Added local warning suppression around fftn/ifftn calls when dtype is float16
- Only suppresses the specific ComplexHalf experimental warning