training_config.yaml

ds_cfg:
  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
  gradient_accumulation_steps: ${gradient_accumulation_steps}
  scheduler:
    type: WarmupCosineLR
    params:
      total_num_steps: null
      warmup_num_steps: null
      warmup_type: linear
  optimizer:
    type: AdamW
    params:
      lr: ${learning_rate}
      betas:
      - 0.9
      - 0.95
      eps: 1.0e-06
      weight_decay: ${weight_decay}
  bf16:
    enabled: false
  zero_optimization:
    stage: 1
    offload_optimizer:
      device: cpu
      pin_memory: true
    stage3_param_persistence_threshold: 100000.0
    stage3_max_live_parameters: 100000000.0
    stage3_prefetch_bucket_size: 100000000.0
    memory_efficient_linear: false
  steps_per_print: 1
  gradient_clipping: 1.0
  prescale_gradients: false
  fp16:
    enabled: true
    auto_cast: false
    loss_scale: 0
    initial_scale_power: 16
    loss_scale_window: 1000
    hysteresis: 2
    consecutive_hysteresis: false
    min_loss_scale: 1
data_path_prefix: null
model_path_prefix: /home/aiops/yangph/checkpoints/glide/checkpoint/experiments/
output_path_prefix: /home/aiops/yangph/checkpoints/glide/checkpoint/
train_file: /home/aiops/yangph/data/raw_data/sailor2_sft_data_1101/sailor2_sft_1119_400k_stage2_only1turn.jsonl
dev_file: null
test_file: null
torch_dtype:
  _target_: general_util.training_utils.return_torch_dtype
  dtype: bfloat16
eos_token: <|endoftext|>
eos_token_id: 151643
tokenizer_init:
  _target_: general_util.tokenization_utils.init_tokenizer
  tokenizer_path: ${model_name_or_path}
  padding_side: right
  pad_token: ${eos_token}
device_map:
  _target_: models.utils.return_single_device_map
model:
  _target_: models.qwen2_glide_du_2.Qwen2Glide.from_pretrained
  gradient_checkpointing: false
  attn_implementation: flash_attention_2
  ignore_mismatched_sizes: true
  torch_dtype: ${torch_dtype}
  pad_token_id: ${eos_token_id}
read_tensor:
  _target_: data.combine_dataset.MultiMappingDataset
  read_fn:
    _target_: data.input_utils.jsonl_read_fn
  aligner:
    _target_: data.input_aligner.add_id_aligner
    id_field: id
  template:
    _target_: data.input_utils.recompose_template
    units:
      chat_prefix: ''
      question: '{user}'
      answer: '{assistant}'
      chat_suffix: ${eos_token}
    compositions:
      prompt: '{question}'
      chosen: '{answer}{chat_suffix}'
  instruction: ''
  index_field: id
  kv_mapping:
    chosen: chosen
    id: index
    prompt: prompt
dist_load_data_barrier: false
extended_vocab: null
collator:
  _target_: data.general_collator.SFTCollator
  tokenizer: ${tokenizer_init}
  max_seq_length: 4096
num_workers: 8
prefetch_factor: 2
model_name: sailor2_8B_sft_500step_hf_1102_longxu.sailor2_5b.glide.A100.tp1dp1.zero1.v1.0.8gpu_du_2.s42/checkpoint-last
model_name_or_path: ${model_path_prefix}/${model_name}/
pretrain: null
resume: latest
dp_size: 8
tp_size: 1
pp_size: 1
wandb_project: sailor2
exp_name: sailor2_8b_sft.glide.A100.tp1dp1.zero1.v1.0.8gpu_du_2.5e-6.s${seed}
exp_notes: null
output_dir: ${output_path_prefix}experiments/${exp_name}
do_train: true
evaluate_during_training: false
do_eval: false
eval_sub_path: checkpoint-*
per_gpu_train_batch_size: 2
per_gpu_eval_batch_size: 2
learning_rate: 5.0e-06
gradient_accumulation_steps: 64
weight_decay: 0.1
adam_epsilon: 1.0e-06
adam_betas: (0.9, 0.98)
total_dataset_len: -1
max_grad_norm: 1.0
num_train_epochs: 2
max_steps: 0
warmup_proportion: 0.1
warmup_steps: 0
optimizer: null
use_nvlamb: null
bit_training: null
logging_steps: 1
save_ds_state: true
save_steps: 200
save_best: false
eval_steps: 200
ddp_eval: true
no_cuda: false
seed: 42
local_rank: 0
fp16: true
fp16_opt_level: O1
fp16_bfloat16: true
prediction_cfg:
  metric: loss
  measure: -1
  best_checkpoint: null
  best_result: null
eval_forward_fn:
  _target_: general_util.evaluator.DefaultForwardFn
post_process:
  _target_: post_processors.dpo.SFTLossOnlyPostProcessor
summary_helper:
  _target_: general_util.tensorboard_helper.WandbWriter
  batch_index_or_keys: null
  outputs_index_or_keys:
    train/loss: loss
    train/large_loss: llm_loss
n_gpu: 1
device: cuda:0
train_batch_size: null
eval_batch_size: null
world_size: 8