seed: 0
run_mode: 'finetune'
output_dir: './output' # path to save checkpoint/strategy
load_checkpoint:  ''
auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model
resume_training: False

# ==== context config ====
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  max_device_memory: "57GB"
  save_graphs: False
  save_graphs_path: "./graph"
  device_id: 0
  jit_config:
    jit_level: "O1"
  memory_optimize_level: "O1"

# ==== model config ====
model:
  model_config:
    type: YiZhaoConfig
    batch_size: 1   # only for incremental infer
    num_layers: 54
    padded_vocab_size: 151552
    hidden_size: 4096
    ffn_hidden_size: 13696
    kv_channels: 128
    num_attention_heads: 32
    seq_length: 8192
    hidden_dropout: 0
    attention_dropout: 0.05
    layernorm_epsilon: 1.e-5
    rmsnorm: True
    apply_residual_connection_post_layernorm: False
    post_layer_norm: True
    add_bias_linear: False
    add_qkv_bias: True
    bias_dropout_fusion: True
    multi_query_attention: True
    multi_query_group_num: 2
    apply_query_key_layer_scaling: True
    attention_softmax_in_fp32: True
    fp32_residual_connection: False
    quantization_bit: 0
    pre_seq_len: None
    prefix_projection: False
    param_init_type: "float32"
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    rotary_dtype: "float32"
    use_past: False
    use_flash_attention: True # when use FlashAttention, seq_length should be multiple of 16
    use_prompt_flash_attention: False
    use_incre_flash_attention: False
    eos_token_id: 151329
    pad_token_id: 151329
    repetition_penalty: 1.0
    max_decode_length: 256
    checkpoint_name_or_path: ""
    offset: [-1, 2, 2, -1]
    top_k: 1
    top_p: 1
    do_sample: True
    # refactor param
    qkv_concat: False
    mlp_concat: False
    use_llama_rope: True
    lm_head_dp: False
    pp_interleave_num: 1
    auto_register: yizhao_config.YiZhaoConfig
  arch:
    type: YiZhaoForCausalLM
    auto_register: yizhao.YiZhaoForCausalLM

trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'YiZhao'

# ==== dataset config ====
train_dataset: &train_dataset
  data_loader:
     type: MindDataset
     dataset_dir:  ''
     shuffle: False
  input_columns: ["input_ids", "labels", "attention_mask", "loss_mask"]
  output_columns: ["input_ids", "labels", "attention_mask", "loss_mask"]
  num_parallel_workers: 8
  python_multiprocessing: False # True: muti-process; False: muti-thread
  drop_remainder: True
  batch_size: 1
  repeat: 1
  numa_enable: False
  prefetch_size: 1
  seed: 0
  gmask_id: 151331
  sop_id: 151333
  eos_id: 151329
train_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *train_dataset

# ==== runner config ====
runner_config:
  epochs: 10
  batch_size: 1
  sink_mode: True
  sink_size: 1

runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense: 1
  use_clip_grad: True

lr_schedule:
  type: MegatronLR #CosineWithWarmUpLR
  learning_rate: 5.e-5
  lr_end: 5.e-6
  warmup_ratio: 0.02
  decay_ratio: 0.9
  total_steps: -1 # -1 means it will load the total steps of the dataset
  auto_register: yizhao_scheduler.MegatronLR

# optimizer
optimizer:
  type: AdamW
  betas: [0.9,0.95]
  eps: 1.e-8
  weight_decay: 0.1

# parallel config
use_parallel: True
parallel:
  gradients_mean: False
  loss_repeated_mean: True
  full_batch: True
  enable_parallel_optimizer: True  # optimizer shard
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64
parallel_config:
  data_parallel: 2
  model_parallel: 1
  pipeline_stage: 4
  micro_batch_num: 8
  vocab_emb_dp: False
  use_seq_parallel: True 
  gradient_aggregation_group: 4
micro_batch_interleave_num: 1

# recompute
recompute_config:
  recompute: True
  select_recompute: [
    'mlp\.activation_func\.mul',
    'mlp\.activation_func\.silu\.silu',
    'mlp\.dense_left\.reshape',
    'mlp\.dense_4h_to_h\.reshape',
    'input_layernorm\.cast',
    'post_attention_layernorm\.cast']
  parallel_optimizer_comm_recompute: False
  mp_comm_recompute: True
  recompute_slice_activation: False