scripts/optimize_nccl.sh

#!/usr/bin/env bash
# NCCL optimization script for distributed training
# Usage: source scripts/optimize_nccl.sh

# NCCL timeout and communication settings
export NCCL_TIMEOUT=3600                    # 1 hour timeout (default 30min)
export NCCL_ASYNC_ERROR_HANDLING=1         # Enable async error handling
export NCCL_MAX_NCHANNELS=4                # Reduce number of channels for better stability
export NCCL_MIN_NCHANNELS=4                # Set minimum channels
export NCCL_BUFFSIZE=4194304               # 4MB buffer size

# Socket communication optimization
export NCCL_SOCKET_IFNAME=eth0             # Use specific network interface
export NCCL_IB_DISABLE=1                   # Disable InfiniBand if not available

# GPU memory and communication optimization
export CUDA_LAUNCH_BLOCKING=0              # Non-blocking CUDA operations
export NCCL_P2P_DISABLE=0                  # Enable peer-to-peer communication
export NCCL_SHM_DISABLE=0                  # Enable shared memory

# Debugging (set to 0 for production)
export NCCL_DEBUG=INFO                     # INFO level debugging
export TORCH_DISTRIBUTED_DEBUG=DETAIL      # Detailed distributed debugging

echo "NCCL optimization settings applied:"
echo "- NCCL_TIMEOUT: $NCCL_TIMEOUT"
echo "- NCCL_ASYNC_ERROR_HANDLING: $NCCL_ASYNC_ERROR_HANDLING"
echo "- NCCL_MAX_NCHANNELS: $NCCL_MAX_NCHANNELS"
echo "- NCCL_DEBUG: $NCCL_DEBUG"
echo ""
echo "To use these settings, run:"
echo "  source scripts/optimize_nccl.sh"
echo "  bash scripts/run_all_ablations.sh --group A"