seed: 0
output_dir: './output' # path to save checkpoint/strategy
load_checkpoint: '' # 必填
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: 'predict'
use_parallel: True

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'yi_34b'

# runner config
runner_config:
  epochs: 2
  batch_size: 32
  sink_mode: True
  sink_size: 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: 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 = 8 for Atlas 800
parallel_config:
  data_parallel: 1
  model_parallel: 4
  pipeline_stage: 1
  use_seq_parallel: False
  micro_batch_num: 4
  vocab_emb_dp: 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

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

# mindspore context init config
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  device_target: "Ascend"
  max_call_depth: 10000
  max_device_memory: "58GB"
  save_graphs: False
  save_graphs_path: "./graph"

# model config
model:
  model_config:
    type: LlamaConfig
    batch_size: 32 # add for increase predict
    seq_length: 16384
    hidden_size: 7168
    num_layers: 60
    num_heads: 56
    max_position_embedding: 4096
    vocab_size: 64000
    intermediate_size: 20480
    multiple_of: 256
    n_kv_heads: 8
    rms_norm_eps: 1.0e-5
    bos_token_id: 6
    eos_token_id: 7
    pad_token_id: 0
    ignore_token_id: -100
    theta: 5000000.0
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "bfloat16"
    param_init_type: "bfloat16"
    use_past: True
    pretrain_seqlen: 4096 # seqlen of the pretrain checkpoint: 2048 for llama and 4096 for llama2
    extend_method: "None" # support "None", "PI", "NTK"
    compute_in_2d: False
    use_flash_attention: True # FA can accelerate training or finetune
    block_size: 128
    num_blocks: 2048
    offset: 0
    use_past_shard: False
    checkpoint_name_or_path: ""
    repetition_penalty: 1.0
    max_decode_length: 4074
    top_k: 40
    top_p: 0.8
    temperature: 0.6
    do_sample: False
    is_dynamic: True
  arch:
    type: LlamaForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    additional_special_tokens: ['<|im_start|>','<|im_end|>','<|im_sep|>']
    unk_token: '<unk>'
    bos_token: '<|startoftext|>'
    eos_token: '<|endoftext|>'
    pad_token: '<unk>'
    type: LlamaTokenizer
    vocab_file: "" # 必填
    add_bos_token: False
    add_eos_token: False
    chat_template: "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}"
  type: LlamaProcessor