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

load_checkpoint: ''
auto_trans_ckpt: False  # If true, auto transform load_checkpoint to load in distributed model

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'llama2_7b'

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

# mindspore context init config
context:
  mode: 0 #0--Graph Mode; 1--Pynative Mode
  max_device_memory: "58GB"
  device_id: 0

# model config
model:
  model_config:
    type: LlamaConfig
    batch_size: 1 # add for increase predict
    seq_length: 4096
    hidden_size: 4096
    num_layers: 32
    num_heads: 32
    vocab_size: 32000
    multiple_of: 256
    rms_norm_eps: 1.0e-5
    bos_token_id: 1
    eos_token_id: 2
    pad_token_id: 0
    ignore_token_id: -100
    compute_dtype: "float16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "float16"
    param_init_type: "float16"
    use_past: True
    scaling_factor: 1.0 # The scale factor of seq length
    extend_method: "None" # support "None", "PI", "NTK"
    use_flash_attention: True # FA can accelerate training or finetune
    block_size: 16
    num_blocks: 1024
    is_dynamic: True
    qkv_concat: False
    offset: 0
    checkpoint_name_or_path: "llama2_7b"
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 3
    top_p: 1
    do_sample: False
  arch:
    type: LlamaForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    unk_token: '<unk>'
    bos_token: '<s>'
    eos_token: '</s>'
    pad_token: '<unk>'
    type: LlamaTokenizer
  type: LlamaProcessor