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: 'predict'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'llama3_1_8b'

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

# eval dataset
eval_dataset: &eval_dataset
  data_loader:
    type: MindDataset
    dataset_dir: ""
    shuffle: False
  input_columns: ["input_ids"]
  num_parallel_workers: 8
  python_multiprocessing: False
  drop_remainder: False
  repeat: 1
  numa_enable: False
  prefetch_size: 1
eval_dataset_task:
  type: CausalLanguageModelDataset
  dataset_config: *eval_dataset

use_parallel: False
# 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 800T A2
parallel_config:
  data_parallel: 1
  model_parallel: 1
  pipeline_stage: 1
  use_seq_parallel: False
  micro_batch_num: 1
  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

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

# model config
model:
  model_config:
    type: LlamaConfig
    batch_size: 1 # add for increase predict
    seq_length: 512
    hidden_size: 4096
    num_layers: 32
    num_heads: 32
    n_kv_heads: 8
    vocab_size: 128256
    intermediate_size: 14336
    rms_norm_eps: 1.0e-5
    bos_token_id: 128000
    eos_token_id: 128001
    pad_token_id: 128002
    ignore_token_id: -100
    max_position_embedding: 131072
    compute_dtype: "float16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "float32"
    param_init_type: "bfloat16"
    use_past: True
    is_dynamic: True
    theta: 500000
    extend_method: "LLAMA3" # support "None", "PI", "NTK", "LLAMA3"
    scaling_factor:
     factor: 8.0
     low_freq_factor: 1.0
     high_freq_factor: 4.0
     original_max_position_embeddings: 8192
    use_flash_attention: True # FA can accelerate training or finetune
    offset: 0
    fine_grain_interleave: 1
    checkpoint_name_or_path: ""
    repetition_penalty: 1
    max_decode_length: 512
    block_size: 16
    num_blocks: 512
    top_k: 3
    top_p: 1
    do_sample: False
    auto_map:
      AutoTokenizer: [ llama3_1_tokenizer.Llama3Tokenizer, null ]
  arch:
    type: LlamaForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    model_max_length: 8192
    vocab_file: "/path/tokenizer.model"
    pad_token: "<|reserved_special_token_0|>"
    type: Llama3Tokenizer
  type: LlamaProcessor

# metric
metric:
  type: PerplexityMetric

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

eval_callbacks:
  - type: ObsMonitor

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

profile: False
profile_start_step: 4
profile_stop_step: 8
init_start_profile: False
profile_communication: False
profile_memory: True
layer_scale: False
layer_decay: 0.65
lr_scale_factor: 256

# aicc
remote_save_url: "Please input obs url on AICC platform."