FlagScale/examples/aquila/conf/train/1_8b.yaml at 42f45a93f5fa9a56753880d203fedf2abe2d16da · lxd-cumt/FlagScale · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
system:
  #skip_iters_range: [10, 20]
  distributed_backend: flagcx
  reset_position_ids: True
  reset_attention_mask: True
  add_qkv_bias: True
  tensor_model_parallel_size: 1
  pipeline_model_parallel_size: 2
  disable_bias_linear: True
  use_distributed_optimizer: True
  precision:
    bf16: True
    min_loss_scale: 1.0
    attention_softmax_in_fp32: True
    accumulate_allreduce_grads_in_fp32: True
  logging:
    log_interval: 1
    log_throughput: True
    tensorboard_log_interval: 1
    wandb_project: ${experiment.exp_name}
    wandb_exp_name: ${experiment.exp_name}
  checkpoint:
    ckpt_format: torch
    save_interval: 5

  hetero:
    enable_hetero: True
    hetero_use_cpu_communication: False
    use_partial_reduce_for_shared_embedding: True
    # mesh format [tp1,cp1,ep1,dp1,pp1,(tp2,cp2...)]
    hetero_pipeline_layer_split: [12,12]
    hetero_process_meshes: [1,1,1,2,1, 1,1,1,2,1]
    hetero_device_types: ["A800", "mlu590"]

    standalone_embedding_stage: False
    hetero_current_device_type: "A800"

model:
  transformer_impl: transformer_engine
  num_layers: 24
  hidden_size: 2048
  num_attention_heads: 16
  seq_length: 4096
  max_position_embeddings: 4096 # only for adding position embeddings
  norm_epsilon: 1e-6
  norm_init_weight: 0.8441
  use_rotary_position_embeddings: true
  no_position_embedding: true
  rotary_base: 1000000
  swiglu: true
  multiple_of: 128
  #hidden_dim_multiplier: 2 # ffn_hidden_size 11008 / 5504
  normalization: RMSNorm
  position_embedding_type: rope
  untie_embeddings_and_output_weights: True
  init_method_std: 0.02
  attention_dropout: 0.0
  hidden_dropout: 0.0
  weight_decay: 0.1
  clip_grad: 1.0
  train_samples: 8192  #8192 #29297664 #120B tokens 100step
  eval_iters: 0
  eval_interval: 1000
  micro_batch_size: 8
  global_batch_size: 1024
  seed: 42

  optimizer:
    weight_decay: 0.1
    adam_beta1: 0.9
    adam_beta2: 0.95
    lr_scheduler:
      lr: 3.0e-4
      min_lr: 3.0e-5
      lr_warmup_samples: 2048000
      lr_wsd_decay_style: cosine


data:
  data_path: xxx # Please replace with your actual data path
  split: 1
  no_mmap_bin_files: true
  tokenizer:
    tokenizer_type: QwenTokenizerFS
    tokenizer_path: xxx xxxx # Please replace with your actual tokenizer path
    vocab_size: xxx # Please replace with your actual vocab size
    make_vocab_size_divisible_by: 64