hails commited on
Commit
fab2ccd
1 Parent(s): 00fc3df

Upload pythia-dedup-125M.yml

Browse files
Files changed (1) hide show
  1. pythia-dedup-125M.yml +107 -0
pythia-dedup-125M.yml ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "pipe-parallel-size": 1,
3
+ "model-parallel-size": 1,
4
+
5
+ # model settings
6
+ "num-layers": 12,
7
+ "hidden-size": 768,
8
+ "num-attention-heads": 12,
9
+ "seq-length": 2048,
10
+ "max-position-embeddings": 2048,
11
+ "pos-emb": "rotary",
12
+ "rotary-pct": 0.25,
13
+ "no-weight-tying": true,
14
+ "gpt-j-residual": true,
15
+ "output-layer-parallelism": "column",
16
+
17
+ "scaled-upper-triang-masked-softmax-fusion": true,
18
+ "bias-gelu-fusion": true,
19
+
20
+ # init methods
21
+ "init_method": "small_init",
22
+ "output_layer_init_method": "wang_init",
23
+
24
+ "optimizer": {
25
+ "type": "Adam",
26
+ "params": {
27
+ "lr": 0.0006,
28
+ "betas": [0.9, 0.95],
29
+ "eps": 1.0e-8,
30
+ }
31
+ },
32
+ "min_lr": 0.00006,
33
+
34
+ "zero_optimization": {
35
+ "stage": 1,
36
+ "allgather_partitions": True,
37
+ "allgather_bucket_size": 500000000,
38
+ "overlap_comm": True,
39
+ "reduce_scatter": True,
40
+ "reduce_bucket_size": 500000000,
41
+ "contiguous_gradients": True,
42
+ "cpu_offload": False
43
+ },
44
+
45
+ "train_micro_batch_size_per_gpu": 16,
46
+ "gas": 2,
47
+ "data-impl": "mmap",
48
+ "num_workers": 1,
49
+
50
+ # activation checkpointing
51
+ "checkpoint-activations": true,
52
+ "checkpoint-num-layers": 1,
53
+ "partition-activations": true,
54
+ "synchronize-each-layer": true,
55
+
56
+ # regularization
57
+ "gradient_clipping": 1.0,
58
+ "weight-decay": 0.1,
59
+ "hidden-dropout": 0,
60
+ "attention-dropout": 0,
61
+
62
+ # precision settings
63
+ "fp16": {
64
+ "fp16": true,
65
+ "enabled": true,
66
+ "loss_scale": 0,
67
+ "loss_scale_window": 1000,
68
+ "initial_scale_power": 12,
69
+ "hysteresis": 2,
70
+ "min_loss_scale": 1,
71
+ },
72
+
73
+ "train-iters": 71500,
74
+ "lr-decay-iters": 71250,
75
+ "distributed-backend": "nccl",
76
+ "lr-decay-style": "cosine",
77
+ "warmup": 0.01,
78
+ "save-interval": 250,
79
+ "eval-interval": 71500,
80
+ "eval-iters": 10,
81
+ #"eval-tasks": ["logiqa", "arc_challenge", "arc_easy", "lambada", "winogrande", "wsc", "piqa", "sciq"],
82
+
83
+ "log-interval": 10,
84
+ "steps_per_print": 10,
85
+ "wall_clock_breakdown": true,
86
+
87
+ "save": "/fsx/hailey/pythia/ckpts/125M_dedup",
88
+ "load": "/fsx/hailey/pythia/ckpts/125M_dedup",
89
+
90
+ "train-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"],
91
+ "valid-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"],
92
+ "test-data-paths": ["/fsx/pile_deduped/pile_0.87_deduped_text_document"],
93
+
94
+ "tokenizer-type": "HFTokenizer",
95
+ "vocab-file": "/fsx/pile/20B_tokenizer.json",
96
+
97
+ "tensorboard-dir": "/fsx/code-fim/FIMlogs/1.3B-AR-Pile-9-6-22-rotary-1MtokBS",
98
+ "log-dir": "/fsx/code-fim/FIMlogs/1.3B-AR-Pile-9-6-22-rotary-1MtokBS",
99
+
100
+ "use_wandb": true,
101
+ "wandb_group": "125M Dedup",
102
+ "wandb_team": "eleutherai",
103
+ "wandb_project": "pythia",
104
+
105
+ "launcher": "openmpi",
106
+ "deepspeed_mpi": true,
107
+ }