multi_gpu_setup.yaml

# Multi-GPU Setup Configuration
# ThemisDB v1.4.0-alpha
# Use case: High-performance training, multi-GPU inference

# Server Configuration
server:
  host: 0.0.0.0
  port: 8080
  workers: 16
  max_connections: 5000
  
  request_timeout: 600
  keepalive_timeout: 120
  
  max_request_size: 500MB
  max_response_size: 500MB

# Multi-GPU Configuration
gpu:
  enabled: true
  
  # Multiple GPUs
  devices:
    - id: 0
      name: "GPU-0-A100"
      memory_limit: 0.9
      compute_capability: 8.0
      
    - id: 1
      name: "GPU-1-A100"
      memory_limit: 0.9
      compute_capability: 8.0
      
    - id: 2
      name: "GPU-2-A100"
      memory_limit: 0.9
      compute_capability: 8.0
      
    - id: 3
      name: "GPU-3-A100"
      memory_limit: 0.9
      compute_capability: 8.0
  
  # Multi-GPU strategy
  multi_gpu:
    enabled: true
    strategy: data_parallel  # Options: data_parallel, model_parallel, pipeline_parallel
    
    # Data parallel settings
    data_parallel:
      gradient_sync: true
      bucket_cap_mb: 25
      
    # Model parallel settings (if strategy: model_parallel)
    model_parallel:
      tensor_parallel_size: 4
      pipeline_parallel_size: 1
      tensor_split: [0.25, 0.25, 0.25, 0.25]  # Equal distribution
  
  # NCCL configuration for multi-GPU communication
  nccl:
    enabled: true
    debug: false
    
    # Communication settings
    p2p_enabled: true  # GPU-to-GPU direct transfers via NVLink
    ib_enabled: false  # Set to true for InfiniBand
    
    # Network settings
    socket_ifname: "eth0"
    net_plugin: ""
    
    # Performance tuning
    buffsize: 2097152
    nthreads: 4
    algo: "Ring"  # Options: Ring, Tree, CollNet
    proto: "Simple"  # Options: Simple, LL, LL128
    
    # Timeouts
    timeout: 1800  # 30 minutes
  
  # GPU settings
  persistence_mode: true
  compute_mode: exclusive_process
  power_limit: 400  # Watts per GPU
  
  # Memory management
  memory_pool:
    enabled: true
    initial_size_mb: 8192
    max_split_size_mb: 1024
  
  vram:
    secure_clear: true
    unified_memory: false
    oom_protection:
      enabled: true
      reserved_memory_mb: 2048

# LLM Configuration
llm:
  enabled: true
  backend: cuda
  
  # Large model for multi-GPU
  model_path: /models/llama-2-70b-chat.gguf
  model_type: llama
  
  # Increased context and batch for multi-GPU
  context_length: 8192
  batch_size: 2048
  n_gpu_layers: -1  # Auto-distribute across all GPUs
  
  # Multi-GPU tensor split
  tensor_split: [0.25, 0.25, 0.25, 0.25]
  
  # Performance
  threads: 32
  use_mmap: true
  use_mlock: true  # Lock in RAM
  
  # Inference settings
  inference:
    temperature: 0.7
    top_p: 0.9
    top_k: 40
    repeat_penalty: 1.1
    
    kv_cache:
      enabled: true
      max_tokens: 16384  # Larger cache for multi-GPU
      block_size: 16

# LoRA Configuration
lora:
  enabled: true
  adapter_path: /adapters
  max_adapters: 16  # More adapters with more GPUs
  gpu_enabled: true
  
  # Multi-GPU LoRA
  multi_gpu_lora:
    enabled: true
    replicate_adapters: true  # Replicate across GPUs
    
  preload_adapters: 
    - finance-adapter-v1
    - medical-adapter-v2
  cache_adapters: true
  
  rank: 16
  alpha: 32
  dropout: 0.05

# Training Configuration
training:
  enabled: true
  
  # Larger batch sizes for multi-GPU
  batch_size: 128  # Effective batch size
  micro_batch_size: 32  # Per-GPU batch size
  gradient_accumulation_steps: 4  # 128 / 32 = 4
  
  # Precision
  precision: bf16  # BF16 better for A100
  mixed_precision:
    enabled: true
  
  # Optimization
  optimizer: adamw
  learning_rate: 5e-4  # Higher LR for larger batch
  weight_decay: 0.01
  
  lr_schedule:
    type: cosine_with_warmup
    warmup_steps: 2000
    min_lr: 5e-5
  
  # Gradient management
  gradient_clipping:
    enabled: true
    max_norm: 1.0
  
  gradient_checkpointing:
    enabled: false  # Disabled with enough VRAM
  
  # Distributed training
  distributed:
    backend: nccl
    find_unused_parameters: false
    
    # Zero Redundancy Optimizer (optional)
    zero:
      enabled: false
      stage: 2  # Options: 1, 2, 3
  
  # Checkpoints
  checkpoint:
    enabled: true
    save_interval: 1000
    max_keep: 5
    output_dir: /data/checkpoints
    compression: true
    async_save: true
    
    # Distributed checkpoint
    distributed_checkpoint: true
    shard_checkpoint: true
    
  validation:
    enabled: true
    interval: 500
    split: 0.05

# Inference Configuration
inference:
  enabled: true
  
  # High-throughput batching
  batch_size: 64
  continuous_batching:
    enabled: true
    max_batch_size: 128
    max_wait_time_ms: 100
  
  # Precision
  precision: fp16
  
  # Generation
  max_tokens: 1024
  stream: true
  
  # Multi-GPU inference
  multi_gpu_inference:
    enabled: true
    load_balancing: round_robin  # Options: round_robin, least_loaded
    
  # KV cache optimization
  kv_cache:
    enabled: true
    max_tokens: 65536
    block_size: 16

# Storage Configuration
storage:
  data_dir: /data/themisdb
  
  # High-performance RocksDB settings
  rocksdb:
    max_open_files: 5000
    max_background_jobs: 16
    write_buffer_size: 256MB
    max_write_buffer_number: 6
    
    # Compression
    compression: zstd
    compression_level: 3
    
    # Cache
    block_cache_size: 8GB
    
    # Write ahead log
    wal_dir: /data/themisdb/wal
    wal_sync: true
    
  # Encryption
  encryption:
    enabled: true
    algorithm: AES-256-GCM
    key_provider: vault

# Network Configuration
network:
  http:
    enabled: true
    port: 8080
    workers: 16
  
  binary:
    enabled: true
    port: 18765
    workers: 16
  
  grpc:
    enabled: true
    port: 50051
  
  # TLS
  tls:
    enabled: true
    min_version: "1.3"
    cert_file: /etc/themisdb/certs/server.crt
    key_file: /etc/themisdb/certs/server.key
    ca_file: /etc/themisdb/certs/ca.crt
    
    # mTLS
    client_auth: required
    client_ca_file: /etc/themisdb/certs/client-ca.crt

# Monitoring Configuration
metrics:
  enabled: true
  port: 4318
  path: /metrics
  
  gpu_metrics_interval: 1s
  training_metrics_interval: 1s
  inference_metrics_interval: 100ms
  
  # Multi-GPU metrics
  per_gpu_metrics: true
  nccl_metrics: true
  
  exporters:
    - type: prometheus
      endpoint: http://localhost:4318/metrics
    
    - type: opentelemetry
      endpoint: http://otel-collector:4317

# Logging Configuration
logging:
  level: info
  format: json
  output: /var/log/themisdb/app.log
  
  rotation:
    max_size: 500MB
    max_age: 30
    max_backups: 10
    compress: true
  
  # Structured logging
  fields:
    service: themisdb
    environment: production
    cluster: multi-gpu-cluster
  
  audit:
    enabled: true
    output: /var/log/themisdb/audit.log
    events:
      - authentication
      - gpu_allocation
      - model_load
      - training_start

# Security Configuration
security:
  authentication:
    enabled: true
    method: mtls
    
    # API keys as fallback
    api_keys:
      enabled: true
      header: X-API-Key
  
  authorization:
    enabled: true
    rbac:
      enabled: true
  
  # GPU access control
  gpu:
    access_control:
      enabled: true
      mode: strict
      per_user_limit: 1  # Max 1 GPU per user
      per_process_limit:
        max_vram_mb: 10240  # 10GB max per process
  
  # Encryption
  encryption:
    enabled: true
    
  # Audit
  audit:
    enabled: true
    comprehensive: true

# Backup Configuration
backup:
  enabled: true
  
  schedule:
    full_backup: "0 1 * * 0"  # Sunday 1 AM
    incremental: "0 1 * * 1-6"  # Daily 1 AM
    checkpoint: "0 */6 * * *"  # Every 6 hours
  
  retention:
    full: 4
    incremental: 14
    checkpoint: 48
  
  destinations:
    - type: local
      path: /backup/themisdb
      
    - type: s3
      bucket: themisdb-backups
      region: us-west-2
      encryption: true

# Resource Limits
limits:
  max_memory_mb: 262144  # 256GB
  max_disk_usage_gb: 2000  # 2TB
  max_concurrent_requests: 1000
  max_queue_size: 10000
  
  # Per-GPU limits
  per_gpu_limits:
    max_concurrent_jobs: 4
    max_memory_allocation: 0.9

# High Availability
ha:
  enabled: false
  # HA configuration for production clusters

# Performance Tuning
performance:
  # Dataloader optimization
  dataloader:
    num_workers: 32
    prefetch_factor: 8
    pin_memory: true
    persistent_workers: true
  
  # I/O optimization
  io:
    async_io: true
    read_ahead_kb: 1024
    
  # Compilation
  compilation:
    enabled: true
    backend: inductor  # PyTorch 2.0 compilation

# Development Settings
development:
  debug_mode: false
  profiling: false
  hot_reload: false

# Example Usage:
# 1. Ensure all 4 GPUs are detected: nvidia-smi
# 2. Copy this file to /etc/themisdb/config.yaml
# 3. Adjust GPU IDs and paths
# 4. Set NCCL environment variables if needed
# 5. Start service: sudo systemctl start themisdb
# 6. Monitor multi-GPU usage: nvidia-smi dmon -s pucvmet