paddlenlp

1
{
2
  "dp_degree": "auto",
3
  "invalid_strategy": [
4
    "stage3_mp*"
5
  ],
6
  "max_search_time": 900,
7
  "max_time_per_task": 300,
8
  "metric_cfg": {
9
    "OptimizationDirection": "Maximize",
10
    "name": "interval_samples_per_second"
11
  },
12
  "micro_batch_size": "auto",
13
  "mode": "LoRA",
14
  "model_cfg": {
15
    "global_batch_size": 8,
16
    "hidden_size": 4096,
17
    "num_attention_heads": 32,
18
    "num_layers": 28,
19
    "vocab_size": 65024
20
  },
21
  "mp_degree": [
22
    1
23
  ],
24
  "need_baseline": true,
25
  "pp_degree": [
26
    1
27
  ],
28
  "run_cmd": {
29
    "gradient_accumulation_steps": [
30
      "./autoconfig/llama7b_lora_params.json",
31
      "gradient_accumulation_steps"
32
    ],
33
    "micro_batch_size": [
34
      "./autoconfig/llama7b_lora_params.json",
35
      "per_device_train_batch_size"
36
    ],
37
    "mp_degree": [
38
      "./autoconfig/llama7b_lora_params.json",
39
      "tensor_parallel_degree"
40
    ],
41
    "pp_degree": [
42
      "./autoconfig/llama7b_lora_params.json",
43
      "pipeline_parallel_degree"
44
    ],
45
    "run_best_stage": {
46
      "autotuner_benchmark": [
47
        "./autoconfig/llama7b_lora_params.json",
48
        "autotuner_benchmark",
49
        0
50
      ]
51
    },
52
    "search_stage": {
53
      "autotuner_benchmark": [
54
        "./autoconfig/llama7b_lora_params.json",
55
        "autotuner_benchmark",
56
        1
57
      ]
58
    },
59
    "sharding_degree": [
60
      "./autoconfig/llama7b_lora_params.json",
61
      "sharding_parallel_degree"
62
    ],
63
    "sharding_stage": [
64
      "./autoconfig/llama7b_lora_params.json",
65
      "sharding",
66
      "stage"
67
    ],
68
    "use_recompute": [
69
      "./autoconfig/llama7b_lora_params.json",
70
      "recompute"
71
    ],
72
    "recompute_granularity": [
73
      "./autoconfig/llama7b_lora_params.json",
74
      "recompute_granularity"
75
    ]
76
  },
77
  "schedule_prior": [
78
    "mp4"
79
  ],
80
  "sharding_degree": "auto",
81
  "sharding_stage": "auto",
82
  "task_limit": 2000,
83
  "use_recompute": "auto",
84
  "recompute_granularity":"auto"
85
}
86