seed: 0
output_dir: './output'
load_checkpoint: ""
src_strategy_path_or_dir: ''
auto_trans_ckpt: False
only_save_strategy: False
resume_training: False
run_mode: 'finetune'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'qwen2-57b-a14b'

do_eval: false
eval_step_interval: -1
eval_epoch_interval: 1

# runner config
runner_config:
  epochs: 2
  batch_size: 1
  sink_mode: True
  sink_size: 1

# optimizer
optimizer:
  type: AdamW
  betas: [0.9,0.95]
  eps: 1.e-8
  weight_decay: 0.01

# moe
moe_config:
  expert_num: 64
  capacity_factor: 1.1
  aux_loss_factor: 0.001
  num_experts_chosen: 8
  routing_policy: "TopkRouterV2"
  enable_sdrop: True
  router_dense_type: "float32"
  shared_expert_num: 8.0
  use_shared_expert_gating: True
  use_fused_ops_topkrouter: True
  norm_topk_prob: False

# lr sechdule
lr_schedule:
  type: CosineWithWarmUpLR
  learning_rate: 1.e-5
  lr_end: 1.e-5
  warmup_ratio: 0.0
  total_steps: -1

# dataset
train_dataset: &train_dataset
  data_loader:
    type: MindDataset
    dataset_dir: ""
    shuffle: False
  input_columns: ["input_ids", "target_ids", "attention_mask"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  batch_size: 4
  repeat: 1
  numa_enable: False
  prefetch_size: 1
train_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *train_dataset


# eval dataset
eval_dataset: &eval_dataset
  data_loader:
    type: MindDataset
    dataset_dir: ""
    shuffle: False
  input_columns: ["input_ids"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: False
  repeat: 1
  numa_enable: False
  prefetch_size: 1
eval_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *eval_dataset

use_parallel: True
# parallel context config
parallel:
  parallel_mode: 1 # 0-data parallel, 1-semi-auto parallel, 2-auto parallel, 3-hybrid parallel
  gradients_mean: False
  enable_alltoall: True
  full_batch: True
  search_mode: "sharding_propagation"
  enable_parallel_optimizer: True
  strategy_ckpt_save_file: "./ckpt_strategy.ckpt"
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64
# default parallel of device num = 8 910A
parallel_config:
  data_parallel: 4
  expert_parallel: 4
  model_parallel: 4
  pipeline_stage: 2
  use_seq_parallel: True
  micro_batch_num: 8
  vocab_emb_dp: False
  gradient_aggregation_group: 4
# when model parallel is greater than 1, we can set micro_batch_interleave_num=2, that may accelerate the train process.
micro_batch_interleave_num: 1

# recompute config
recompute_config:
  recompute: True
  select_recompute: False
  parallel_optimizer_comm_recompute: False
  mp_comm_recompute: True
  recompute_slice_activation: True

# callbacks
callbacks:
 - type: MFLossMonitor
 - type: ObsMonitor

context:
  jit_config:
    jit_level: "O1"
  memory_optimize_level: "O1"
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  enable_graph_kernel: False
  max_call_depth: 10000
  max_device_memory: "58GB"
  save_graphs: False
  save_graphs_path: "./graph"
  device_id: 0

# model config
model:
  model_config:
    type: LlamaConfig
    batch_size: 1 # add for increase predict
    seq_length: 32768
    max_position_embedding: 32768
    hidden_size: 3584
    intermediate_size: 2560
    theta: 1000000
    num_layers: 28
    num_heads: 28
    vocab_size: 151936
    n_kv_heads: 4
    rms_norm_eps: 1.0e-6
    bos_token_id: 151643
    pad_token_id: 151643
    eos_token_id: [151645, 151643]
    ignore_token_id: -100
    qkv_has_bias: True
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float16"
    rotary_dtype: "float32"
    param_init_type: "float32"
    use_past: False
    extend_method: "None" # support "None", "PI", "NTK"
    use_flash_attention: True
    use_past_shard: False
    checkpoint_name_or_path: ""
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 3
    top_p: 1
    do_sample: False
    is_dynamic: False
    use_attn_mask_compression: True
  arch:
    type: LlamaForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    model_max_length: 32768
    vocab_file: ""
    merges_file: ""
    unk_token: "<|endoftext|>"
    eos_token: "<|endoftext|>"
    pad_token: "<|endoftext|>"
    type: Qwen2Tokenizer
  type: Qwen2Processor

# metric
metric:
  type: PerplexityMetric

# wrapper cell config
runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense:
    type: DynamicLossScaleUpdateCell
    loss_scale_value: 4096
    scale_factor: 2
    scale_window: 1000
  use_clip_grad: True

eval_callbacks:
  - type: ObsMonitor

auto_tune: False
filepath_prefix: './autotune'
autotune_per_step: 10

profile: False
profile_start_step: 2
profile_stop_step: 4
init_start_profile: False
profile_communication: False
profile_memory: True
layer_scale: False
layer_decay: 0.65
lr_scale_factor: 256

# aicc
remote_save_url: ""