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
run_mode: 'finetune'

use_parallel: True

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'qwen_14b'

# dataset
train_dataset: &train_dataset
  data_loader:
    type: MindDataset
    dataset_dir: ""
    shuffle: True
  input_columns: ["input_ids", "labels", "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

# runner config
runner_config:
  epochs: 5
  batch_size: 1
  sink_mode: True
  sink_size: 2

runner_wrapper:
  type: MFTrainOneStepCell
  scale_sense:
    type: DynamicLossScaleUpdateCell
    loss_scale_value: 1024
    scale_factor: 2
    scale_window: 1000
  use_clip_grad: True

# optimizer
optimizer:
  type: FP32StateAdamWeightDecay
  beta1: 0.9
  beta2: 0.95
  eps: 1.e-8
  weight_decay: 0.1

# lr schedule
lr_schedule:
  type: CosineWithWarmUpLR
  learning_rate: 1.e-5
  warmup_ratio: 0.01
  total_steps: -1 # -1 means it will load the total steps of the dataset

# callbacks
callbacks:
  - type: MFLossMonitor
  - type: CheckpointMonitor
    prefix: "qwen"
    save_checkpoint_steps: 10000
    keep_checkpoint_max: 3
    integrated_save: False
    async_save: False
  - type: ObsMonitor

# default parallel of device num = 8 for Atlas 800T A2
parallel_config:
  data_parallel: 8
  model_parallel: 1
  pipeline_stage: 1
  micro_batch_num: 1
  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

model:
  model_config:
    type: QwenConfig
    batch_size: 1
    seq_length: 2048
    hidden_size: 5120
    num_layers: 40
    num_heads: 40
    vocab_size: 152064
    intermediate_size: 13696
    rms_norm_eps: 1.0e-6
    emb_dropout_prob: 0.0
    eos_token_id: 151643
    pad_token_id: 151643
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float16"
    rotary_dtype: "float16"
    param_init_type: "float32"
    use_past: True
    use_flash_attention: True
    block_size: 32
    num_blocks: 128
    offset: 0
    checkpoint_name_or_path: "/path/qwen_14b_base.ckpt"
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 0
    top_p: 0.8
    do_sample: False

    # configuration items copied from Qwen
    rotary_pct: 1.0
    rotary_emb_base: 10000
    kv_channels: 128

  arch:
    type: QwenForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    vocab_file: "/path/qwen.tiktoken"
    pad_token: "<|endoftext|>"
    type: QwenTokenizer
  type: QwenProcessor

# mindspore context init config
context:
  jit_config:
    jit_level: "O2"
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  enable_graph_kernel: False
  ascend_config:
    precision_mode: "must_keep_origin_dtype"
  max_call_depth: 10000
  max_device_memory: "59GB"
  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: True
  strategy_ckpt_config:
    save_file: "./ckpt_strategy.ckpt"
    only_trainable_params: False
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64