#!/bin/bash
# Auto-select available GPUs based on memory usage
# Usage: source select_gpus.sh
# Returns: CUDA_VISIBLE_DEVICES and GPU_LIST for training

MEMORY_THRESHOLD=20  # Skip GPUs with >20% memory usage
MAX_GPUS=1           # Maximum number of GPUs to use

# Get GPU info: index, memory_used, memory_total
gpu_info=$(nvidia-smi --query-gpu=index,memory.used,memory.total --format=csv,noheader,nounits)

available_gpus=()
gpu_list=()

while IFS=, read -r idx mem_used mem_total; do
    # Trim whitespace
    idx=$(echo $idx | xargs)
    mem_used=$(echo $mem_used | xargs)
    mem_total=$(echo $mem_total | xargs)

    # Calculate memory usage percentage
    mem_percent=$(awk "BEGIN {printf \"%.2f\", ($mem_used/$mem_total)*100}")

    # Check if GPU is available (memory usage < threshold)
    if (( $(awk "BEGIN {print ($mem_percent < $MEMORY_THRESHOLD)}") )); then
        available_gpus+=($idx)
        gpu_list+=($idx)
        echo "GPU $idx: ${mem_percent}% memory used - AVAILABLE" >&2
    else
        echo "GPU $idx: ${mem_percent}% memory used - BUSY (skipped)" >&2
    fi

    # Stop if we have enough GPUs
    if [ ${#available_gpus[@]} -ge $MAX_GPUS ]; then
        break
    fi
done <<< "$gpu_info"

# Export CUDA_VISIBLE_DEVICES
if [ ${#available_gpus[@]} -eq 0 ]; then
    echo "ERROR: No available GPUs found!" >&2
    exit 1
fi

export CUDA_VISIBLE_DEVICES=$(IFS=,; echo "${available_gpus[*]}")
export GPU_LIST="[$(IFS=,; echo "${gpu_list[*]}")]"
export NUM_GPUS=${#available_gpus[@]}

echo "" >&2
echo "Selected GPUs: $CUDA_VISIBLE_DEVICES" >&2
echo "GPU List for training: $GPU_LIST" >&2
echo "Number of GPUs: $NUM_GPUS" >&2
echo "" >&2

# Print variables for sourcing
echo "export CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES"
echo "export GPU_LIST='$GPU_LIST'"
echo "export NUM_GPUS=$NUM_GPUS"