File size: 3,096 Bytes
9dd9a2a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
119
120
121
122
123
124
aws_output_bucket: s3://panda-us-west-2/experiments/llama2.13b.wudao.sft.combine.legal.v1.0.seq2k.w16.adamw.NA100.0921.ds
data_dir: null
dist_load_data_barrier: false
train_file: data/files/tmp.json
dev_file: null
test_file: null
model:
  _target_: models.llama.LlamaForConditionalGeneration.from_pretrained
  use_peft: false
  gradient_checkpointing: true
  enable_flash_attention: true
  flash_attention_vanilla_torch: true
read_tensor:
  _target_: data.collators.zh_instruct.TextDatasetUnifyV3
  pair_file_list: null
extended_vocab: null
collator:
  _target_: data.collators.flan.FlanCollatorOverCollator
  collator: null
  max_seq_length: 2048
  tokenizer: ${model_name_or_path}
  decoder_only: true
  padding: longest
  padding_side: right
num_workers: 4
prefetch_factor: 2
do_preprocess: false
model_name_or_path: /tmp/llama2.13b.wudao.sft.combine.v1.0.seq2k.w16.adamw.NA100.0803.ds/checkpoint-1750
pretrain: null
exp_name: llama2.13b.wudao.sft.combine.legal.v1.0.seq2k.w16.adamw.NA100.0921.ds
exp_notes: null
output_dir: /tmp/${exp_name}
resume: null
do_train: true
evaluate_during_training: false
do_eval: false
eval_sub_path: checkpoint-*
per_gpu_train_batch_size: 2
per_gpu_eval_batch_size: 1
learning_rate: 1.0e-06
gradient_accumulation_steps: 8
weight_decay: 0.01
adam_epsilon: 1.0e-06
adam_betas: (0.9, 0.99)
max_grad_norm: 1.0
num_train_epochs: 1
total_dataset_len: -1
max_steps: 0
warmup_proportion: 0
warmup_steps: 0
optimizer: null
use_nvlamb: null
bit_training: null
logging_steps: 1
save_best: false
save_steps: 250
eval_steps: 250
ddp_eval: true
no_cuda: false
seed: 42
local_rank: 0
fp16: true
fp16_opt_level: O1
fp16_bfloat16: true
prediction_cfg:
  metric: acc
  measure: 1
  best_checkpoint: null
  best_result: null
eval_forward_fn:
  _target_: general_util.evaluator.DiscriminatorForwardFn
post_process: null
fairscale_config:
  _target_: general_util.fsdp_utils.default_initialize
  fp16: ${fp16}
  move_grads_to_cpu: false
  move_params_to_cpu: false
  flatten_parameters: false
with_lightseq: false
load_lr_scheduler_states: false
ds_cfg:
  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
  gradient_accumulation_steps: ${gradient_accumulation_steps}
  optimizer:
    type: AdamW
    params:
      lr: ${learning_rate}
      betas:
      - 0.9
      - 0.96
      eps: ${adam_epsilon}
      weight_decay: ${weight_decay}
  scheduler:
    type: WarmupDecayLR
    params:
      total_num_steps: 1474
      warmup_max_lr: ${learning_rate}
      warmup_num_steps: 0
      warmup_type: linear
  gradient_clipping: ${max_grad_norm}
  bf16:
    enabled: ${fp16}
  zero_optimization:
    stage: 1
    contiguous_gradients: true
    overlap_comm: true
    reduce_scatter: true
    reduce_bucket_size: 500000000.0
    allgather_bucket_size: 500000000.0
    offload_optimizer:
      device: cpu
      pin_memory: true
  steps_per_print: 1
summary_helper:
  _target_: general_util.tensorboard_helper.WandbWriter
  batch_index_or_keys: null
  outputs_index_or_keys: null
n_gpu: 1
device: cuda:0
train_batch_size: 2
eval_batch_size: null
world_size: 16
world_rank: null