seed: 0
output_dir: './output' # path to save checkpoint/strategy
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
resume_training: False
use_parallel: True
run_mode: 'predict'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'deepseekV2' 

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

# mindspore context init config
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  max_call_depth: 10000
  max_device_memory: "54GB"
  save_graphs: False
  save_graphs_path: "./graph"
  device_id: 0

# 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: False
  full_batch: True
  search_mode: "sharding_propagation"
  enable_parallel_optimizer: False
  strategy_ckpt_save_file: "./ckpt_strategy.ckpt"
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64

# default parallel of device num = 16 for Atlas 800
parallel_config:
  data_parallel: 1
  model_parallel: 16
  pipeline_stage: 1
  expert_parallel: 1
  micro_batch_num: 20
  vocab_emb_dp: True
  use_seq_parallel: 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

# model config
model:
  model_config:
    type: DeepseekV2Config
    auto_register: deepseek2_config.DeepseekV2Config
    batch_size: 1 # add for increase predict
    seq_length: 8192
    hidden_size: 5120
    num_layers: 60
    num_heads: 128
    max_position_embeddings: 163840
    intermediate_size: 12288
    kv_lora_rank: 512
    q_lora_rank: 1536
    qk_rope_head_dim: 64
    v_head_dim: 128
    qk_nope_head_dim: 128
    vocab_size: 102400
    multiple_of: 256
    rms_norm_eps: 1.0e-6
    bos_token_id: 100000
    eos_token_id: 100001
    pad_token_id: 100001
    ignore_token_id: -100
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "float32"
    scaling_factor:
      beta_fast: 32.0
      beta_slow: 1.0
      factor: 40.0
      mscale: 0.707
      mscale_all_dim: 0.707
      original_max_position_embeddings: 4096
    param_init_type: "bfloat16"
    use_past: False
    extend_method: "YARN"
    use_flash_attention: False
    offset: 0
    checkpoint_name_or_path: ""
    repetition_penalty: 1
    max_decode_length: 1024
    top_k: 5
    top_p: 1
    theta: 10000.0
    do_sample: False
  arch:
    type: DeepseekV2ForCausalLM
    auto_register: deepseek2_model.DeepseekV2ForCausalLM

#moe
moe_config:
  expert_num: 160
  # capacity_factor should be large enable to ensure dropless routing
  capacity_factor: 160.0
  aux_loss_factor: 0.05
  num_experts_chosen: 6
  topk_method: "group_limited_greedy"
  routing_policy: "TopkRouterV2"
  enable_sdrop: True
  shared_expert_num: 2
  routed_scaling_factor: 16.0
  norm_topk_prob: False
  first_k_dense_replace: 1
  moe_intermediate_size: 1536 # moe_layer hidden_layer
  # greedy_group_limited strategy, select topk_group from n_group
  topk_group: 3
  n_group: 8
  aux_loss_factors: []
  aux_loss_types: []
  z_loss_factor: 0.0

processor:
  return_tensors: ms
  tokenizer:
    unk_token: '<unk>'
    bos_token: '<｜begin▁of▁sentence｜>'
    eos_token: '<｜end▁of▁sentence｜>'
    pad_token: '<｜end▁of▁sentence｜>'
    type: LlamaTokenizerFast
    vocab_file: './path/tokenizer.model'
    tokenizer_file: './path/tokenizer.json'
  type: LlamaProcessor

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

profile: False
profile_start_step: 1
profile_stop_step: 10
init_start_profile: False
profile_communication: False
profile_memory: True
layer_scale: False
layer_decay: 0.65
lr_scale_factor: 256