# accelerate_config.yaml - 多GPU训练配置

compute_environment: LOCAL_MACHINE
distributed_type: MULTI_GPU
downcast_bf16: 'no'
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: fp16
num_machines: 1
num_processes: 4  # 根据GPU数量调整
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

# RLHF特定设置
gradient_accumulation_steps: 8
gradient_clipping: 1.0
learning_rate: 1e-5
dataloader_drop_last: true