FlagScale/examples/deepseek_v3/conf/train/11b_a2b.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
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
system:
  no_shared_fs: ${experiment.runner.no_shared_fs}
  num_workers: 2
  tensor_model_parallel_size: 1
  pipeline_model_parallel_size: 4
  expert_model_parallel_size: 2
  context_parallel_size: 1
  disable_bias_linear: true
  reset_position_ids: True
  reset_attention_mask: True
  qk_layernorm: true
  sequence_parallel: true
  use_distributed_optimizer: true
  # overlap_grad_reduce: true
  # overlap_param_gather: true
  use_dualpipev: true
  moe_fb_overlap: true
  precision:
    bf16: true
    attention_softmax_in_fp32: true
    accumulate_allreduce_grads_in_fp32: true
  logging:
    log_interval: 1
    tensorboard_log_interval: 1
    wandb_project: ${experiment.exp_name}
    wandb_exp_name: ${experiment.exp_name}
    log_timers_to_tensorboard: true
    log_validation_ppl_to_tensorboard: true
    log_throughput: true
    log_params_norm: true
    log_num_zeros_in_grad: true
    log_memory_to_tensorboard: true
  checkpoint:
    save_interval: ${experiment.save_steps}
    load: ${experiment.load}
    ckpt_format: ${experiment.ckpt_format}

model:
  transformer_impl: transformer_engine
  num_layers: 16
  hidden_size: 2048
  num_attention_heads: 16
  num_query_groups: 16 # num_key_value_heads
  seq_length: 4096
  max_position_embeddings: 4096
  norm_epsilon: 1e-6
  use_rotary_position_embeddings: true
  rotary_base: 1000000
  swiglu: true
  normalization: RMSNorm
  init_method_std: 0.02
  attention_dropout: 0.0
  hidden_dropout: 0.0
  clip_grad: 1.0
  position_embedding_type: rope
  untie_embeddings_and_output_weights: true
  no_position_embedding: true
  no_rope_fusion: true

  # mla args ==================
  multi_latent_attention: true
  kv_lora_rank: 512
  qk_head_dim: 128
  qk_pos_emb_head_dim: 64
  v_head_dim: 128

  # moe args ===================
  ffn_hidden_size: 11264
  moe_ffn_hidden_size: 1408
  moe_grouped_gemm: true
  moe_shared_expert_intermediate_size: 2816
  num_experts: 64
  moe_router_load_balancing_type: "seq_aux_loss"
  moe_router_score_function: sigmoid
  moe_router_enable_expert_bias: true
  moe_router_bias_update_rate: 0.001
  moe_aux_loss_coeff: 0.02
  moe_layer_freq: "[0]+[1]*15"
  # node limited routing
  moe_router_num_groups: 1
  moe_router_group_topk: 1
  moe_router_topk: 6
  moe_router_topk_scaling_factor: 2.446
  moe_token_dispatcher_type: "alltoall"
  moe_shared_expert_overlap: false

  # training
  seed: ${experiment.seed}
  finetune: false
  micro_batch_size: 1
  global_batch_size: 256
  eval_iters: 0
  eval_interval: 1000
  train_samples: 244141056 #1T #29297664 #120B tokens

  optimizer:
    clip_grad: 1.0
    weight_decay: 0.1
    adam_beta1: 0.9
    adam_beta2: 0.95
    lr_scheduler:
      lr: 3.0e-3
      min_lr: 3.0e-4
      lr_warmup_fraction: 0.01
      lr_decay_style: WSD
      lr_wsd_decay_style: cosine
      lr_wsd_decay_samples: 2048


data:
  data_path: /path
  split: 1
  no_mmap_bin_files: true
  tokenizer:
    tokenizer_type: QwenTokenizerFS
    tokenizer_path: /path
    vocab_size: 151851
    make_vocab_size_divisible_by: 64