seed: 0
output_dir: './output' # path to save checkpoint/strategy
load_checkpoint: ''
load_ckpt_format: ''  # format of checkpoint files
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: 'train'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'deepseekV3'

# 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

# lr schedule
lr_schedule:
  type: CosineWithWarmUpLR
  learning_rate: 5.e-6
  lr_end: 1.e-6
  warmup_ratio: 0.02
  total_steps: -1 # -1 means it will load the total steps of the dataset

# dataset
train_dataset: &train_dataset
  data_loader:
    type: MindDataset
    dataset_dir: "./dataset"
    shuffle: False
  input_columns: ["input_ids", "labels"]  # "input_ids" , labels are used in instruction finetune.
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: True
  repeat: 1
  numa_enable: False
  prefetch_size: 1
train_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *train_dataset

# mindspore context init config
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  max_call_depth: 10000
  max_device_memory: "55GB"
  save_graphs: False
  save_graphs_path: "./graph"
  jit_config:
    jit_level: "O1"

# parallel config for device num = 1024
parallel_config:
  data_parallel: 16
  model_parallel: 4
  pipeline_stage: 16
  expert_parallel: 8
  micro_batch_num: &micro_batch_num 32
  vocab_emb_dp: True
  use_seq_parallel: True
  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

# 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_config:
    save_file: "./ckpt_strategy.ckpt"
    only_trainable_params: False
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64

# recompute config
recompute_config:
  recompute: [3, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 3, 2, 0]
  select_recompute: False
  parallel_optimizer_comm_recompute: True
  mp_comm_recompute: True
  recompute_slice_activation: True

# model config
model:
  model_config:
    type: DeepseekV3Config
    auto_register: deepseek3_config.DeepseekV3Config
    batch_size: 1 # add for increase predict
    seq_length: 4096
    hidden_size:  7168
    num_layers: &num_layers 61
    num_heads: 128
    max_position_embeddings: 4096
    intermediate_size: 18432
    kv_lora_rank: 512
    n_kv_heads: 128
    q_lora_rank: 1536
    qk_rope_head_dim: 64
    v_head_dim: 128
    qk_nope_head_dim: 128
    vocab_size: 129280
    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"
    router_dense_type: "float32"
    param_init_type: "float32"
    qkv_concat: False
    use_past: False
    extend_method: "None"
    use_flash_attention: True
    use_fused_swiglu: True
    use_fused_rope: True
    offset: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, -1]
    checkpoint_name_or_path: ""
    theta: 10000.0
    return_extra_loss: True
    mtp_depth: &mtp_depth 1
    mtp_loss_factor: 0.3
  arch:
    type: DeepseekV3ForCausalLM
    auto_register: deepseek3.DeepseekV3ForCausalLM

#moe
moe_config:
  expert_num: &expert_num 256
  expert_group_size: 8
  capacity_factor: 1.5
  aux_loss_factor: 0.05
  num_experts_chosen: 8
  routing_policy: "TopkRouterV2"
  balance_via_topk_bias: &balance_via_topk_bias True
  topk_bias_update_rate: &topk_bias_update_rate 0.001
  use_fused_ops_topkrouter: True
  shared_expert_num: 1
  routed_scaling_factor: 2.5
  norm_topk_prob: True
  first_k_dense_replace: 3
  moe_intermediate_size: 2048
  aux_loss_factors: [0.0001]
  aux_loss_types: ["expert"]
  z_loss_factor: 0.0
  expert_model_parallel: 1
  use_gating_sigmoid: True
  use_gmm: False

# callbacks
callbacks:
  - type: MFLossMonitor
    per_print_times: 1
  # balance topk bias with callback
  - type: TopkBiasBalanceCallback
    balance_via_topk_bias: *balance_via_topk_bias
    topk_bias_update_rate: *topk_bias_update_rate
    num_layers: *num_layers
    mtp_depth: *mtp_depth
    expert_num: *expert_num
    micro_batch_num: *micro_batch_num
  - type: CheckpointMonitor
    prefix: "deepseekv3"
    save_checkpoint_steps: 1000
    keep_checkpoint_max: 5
    integrated_save: False
    async_save: False
    checkpoint_format: "safetensors"

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

profile: False
profile_start_step: 1
profile_stop_step: 10
init_start_profile: False
profile_communication: False
profile_memory: True