seed: 0
output_dir: './output'
load_checkpoint: ''
src_strategy_path_or_dir: ''
auto_trans_ckpt: False # If true, auto transform load_checkpoint to load in distributed model
only_save_strategy: False
transform_process_num: 16
resume_training: False
ignore_data_skip: False
run_mode: 'train'
load_ckpt_format: 'ckpt' # recommend use 'safetensors'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'telechat_7b_moe'

# runner config
runner_config:
  epochs: 10
  batch_size: 1
  sink_mode: True
  sink_size: 2
  gradient_accumulation_steps: 1

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

# moe
moe_config:
  expert_num: 16
  capacity_factor: 1.2
  aux_loss_factors: [0.01]
  aux_loss_types: ["expert"]
  num_experts_chosen: 4
  routing_policy: "TopkRouterV2"
  enable_sdrop: False
  use_fused_ops_topkrouter: True
  router_dense_type: "float32"
  group_wise_a2a: False
  shared_expert_num: 0
  comp_comm_parallel: False
  comp_comm_parallel_degree: 2
  balance_via_topk_bias: False
  topk_bias_update_rate: 0
  ep_extend_tp: True

# lr sechdule
lr_schedule:
  type: CosineWithWarmUpLR
  learning_rate: 3.0e-4
  lr_end: 1.5e-5
  warmup_steps: 2000
  total_steps: 1000000 # -1 means it will load the total steps of the dataset

# dataset
train_dataset: &train_dataset
  data_loader:
    type: MindDataset
    dataset_dir: ""
    shuffle: True
  input_columns: ["input_ids", "labels"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  numa_enable: False
  prefetch_size: 1
train_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *train_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
    optimizer_weight_shard_size: 2
    parallel_optimizer_threshold: 64

# default parallel of device num = 8 910B
parallel_config:
  data_parallel: 128
  expert_parallel: 2
  model_parallel: 4
  pipeline_stage: 2
  use_seq_parallel: True
  micro_batch_num: 4
  vocab_emb_dp: False
  gradient_aggregation_group: 4
# check_for_nan_in_loss_and_grad: True
# 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: [3,0]
  select_recompute: False
  parallel_optimizer_comm_recompute: False
  mp_comm_recompute: True
  recompute_slice_activation: True

# callbacks
callbacks:
  - type: MFLossMonitor
  - type: CheckpointMonitor
    prefix: "telechat_7b_moe"
    checkpoint_format: "ckpt" # recommend use 'safetensors'
    save_checkpoint_steps: 1600 #1800
    keep_checkpoint_max: 4
    integrated_save: False

# mindspore context init config
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  max_call_depth: 10000
  memory_optimize_level: "O1"
  max_device_memory: "53GB"
  save_graphs: False
  save_graphs_path: "./graph"
  jit_config:
    jit_level: O1
  mempool_block_size: "53GB"
  device_id: 0
  ascend_config:
    parallel_speed_up_json_path: "./parallel_speed_up.json"

# model config
model:
  model_config:
    type: TelechatConfig
    auto_register: telechat_config.TelechatConfig
    batch_size: 1 # add for increase predict
    seq_length: 8192
    hidden_size: 4096
    intermediate_size: 6144
    num_layers: 30
    num_heads: 32
    n_kv_heads: 32
    vocab_size: 131072
    rms_norm_eps: 1.0e-5
    bos_token_id: 1
    eos_token_id: 2
    pad_token_id: 3
    ignore_token_id: -100
    sigma: 0.0048
    mean: 0.0
    hidden_dropout_prob: 0.0
    attention_dropout_prob: 0.0
    embed_dropout_prob: 0.0
    res_dtype: "float32"
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "float32"
    param_init_type: "float32"
    use_past: False
    pretrain_seqlen: 8192 # seqlen of the pretrain checkpoint
    extend_method: "None" # support "None", "PI", "NTK"
    use_flash_attention: True # FA can accelerate training or finetune
    offset: 0
    use_past_shard: False
    checkpoint_name_or_path: "telechat_7b_moe"
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 3
    top_p: 1
    do_sample: False
  arch:
    type: TelechatForCausalLM
    auto_register: telechat.TelechatForCausalLM

# wrapper cell config
runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense: 1.0
  use_clip_grad: True

profile: False
profile_start_step: 200
profile_stop_step: 202
init_start_profile: False
profile_communication: True
profile_memory: False
layer_scale: False
layer_decay: 0.65
lr_scale_factor: 256