compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
  deepspeed_multinode_launcher: standard
  zero3_save_16bit_model: true
  zero_optimization:
    stage: 3
    offload_optimizer:
      device: cpu
      pin_memory: true
      buffer_count: 4
    offload_param:
      device: cpu
      pin_memory: true
      buffer_count: 4
    model_parallel:
      enable: true
      tensor_parallel_size: 4
      pipeline_parallel_size: 1
    reduce_bucket_size: 2e5
    stage3_prefetch_bucket_size: 0.1e6  # Reduced further
    stage3_max_live_parameters: 1e6
    stage3_max_reuse_distance: 1e5
    stage3_max_live_gradients: 1e6
    stage3_param_persistence_threshold: 1e4
    sub_group_size: 1e9
  
  activation_checkpointing:
    partition_activations: true
    contiguous_memory_optimization: true
    cpu_checkpointing: true
  
  bf16:
    enabled: false
  fp16:
    enabled: true

  optimizer:
    type: AdamW
    params:
      eight_bit: true
      lr: auto
      betas: auto
      eps: auto
      weight_decay: auto
  
  scheduler:
    type: WarmupDecayLR
    params:
      num_warmup_steps: auto
      num_training_steps: auto

  communication_data_type: fp16
  communication_bucket_size: 500000000
  dist_backend: nccl
  train_micro_batch_size_per_gpu: auto
  gradient_clipping: 1.0

distributed_type: DEEPSPEED
downcast_bf16: auto
machine_rank: 0
main_training_function: main
mixed_precision: bf16
rdzv_backend: static
same_network: true
use_cpu: auto