seed: 0
output_dir: './output'
load_checkpoint: ''
src_strategy_path_or_dir: ''
auto_trans_ckpt: False
only_save_strategy: False
resume_training: False
run_mode: 'predict'

# trainer config
trainer:
  type: CausalLanguageModelingTrainer
  model_name: 'telechat_35b'

use_parallel: True
# 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"
  strategy_ckpt_save_file: "./ckpt_strategy.ckpt"
  parallel_optimizer_config:
    gradient_accumulation_shard: False
    parallel_optimizer_threshold: 64
parallel_config:
  data_parallel: 1
  model_parallel: 2
  pipeline_stage: 1
  use_seq_parallel: False
  micro_batch_num: 1
  vocab_emb_dp: False
  gradient_aggregation_group: 4

# 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"
  device_id: 0
  mempool_block_size: "58GB"
  affinity_cpu_list: "None"

# model config
model:
  model_config:
    type: TelechatConfig
    auto_register: telechat_config.TelechatConfig
    batch_size: 1 # add for increase predict
    net_name: "telechat3_36b"
    seq_length: 8192
    hidden_size: 6144
    num_layers: 64
    layers_group: 1
    num_heads: 48
    n_kv_heads: 8
    embedding_size: 512
    vocab_size: 131072
    eod_reset: False
    out_proj_has_bias: False
    rms_norm_eps: 1.0e-5
    bos_token_id: 1
    eos_token_id: 2
    pad_token_id: 3
    ignore_token_id: -100
    embed_dropout_prob: 0.
    hidden_dropout_prob: 0.
    attention_dropout_prob: 0.
    intermediate_size: 24576
    res_dtype: "bfloat16"
    compute_dtype: "bfloat16"
    layernorm_compute_type: "float32"
    softmax_compute_type: "float32"
    rotary_dtype: "bfloat16"
    param_init_type: "bfloat16"
    use_past: True
    pretrain_seqlen: 8192 # seqlen of the pretrain checkpoint
    extend_method: "None" # support "None", "PI", "NTK"
    use_flash_attention: True # FA can accelerate training or finetune
    block_size: 16
    num_blocks: 512
    is_dynamic: True
    use_past_shard: False
    repetition_penalty: 1
    max_decode_length: 512
    top_k: 3
    top_p: 1
    do_sample: False
    auto_map:
      AutoModel: telechat.TelechatForCausalLM
      AutoConfig: telechat_config.TelechatConfig
      AutoTokenizer: [telechat_tokenizer.TelechatTokenizer, null]
  arch:
    type: TelechatForCausalLM
    auto_register: telechat.TelechatForCausalLM

processor:
  return_tensors: ms
  tokenizer:
    vocab_file: ""
    unk_token: '<unk>'
    bos_token: '<_start>'
    eos_token: '<_end>'
    pad_token: '<_pad>'
    type: TelechatTokenizer
    auto_register: telechat_tokenizer.TelechatTokenizer

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

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

profile: False
profile_start_step: 1
profile_stop_step: 10
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."