ASzecsenyi commited on
Commit
c2c81ad
·
verified ·
1 Parent(s): d12487c

Upload moe_2x512_1qvna0li/meta_000092.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. moe_2x512_1qvna0li/meta_000092.json +146 -0
moe_2x512_1qvna0li/meta_000092.json ADDED
@@ -0,0 +1,146 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 92,
3
+ "val_bpb": 2.9215333618334247,
4
+ "model_config": {
5
+ "run": "moe_2x512",
6
+ "wandb_group": null,
7
+ "device_type": "cuda",
8
+ "config": "moe20m",
9
+ "depth": 2,
10
+ "aspect_ratio": 64,
11
+ "model_dim": 512,
12
+ "head_dim": 128,
13
+ "sequence_len": 1024,
14
+ "vocab_size": 265,
15
+ "rva_blocks": [],
16
+ "recurrent_vocab_sizes": [],
17
+ "kla_blocks": [],
18
+ "mamba_blocks": [],
19
+ "gdn_blocks": [],
20
+ "gla_blocks": [],
21
+ "moe_blocks": [
22
+ -1
23
+ ],
24
+ "d_state": 16,
25
+ "mamba_params": true,
26
+ "kla_kernel": true,
27
+ "mimo_rank": 1,
28
+ "skip_around_kla": true,
29
+ "decoder_mlp": false,
30
+ "use_reparametrisation_trick": true,
31
+ "num_experts": 4,
32
+ "moe_top_k": 2,
33
+ "moe_capacity_factor": 1.0,
34
+ "moe_aux_loss_weight": 0.01,
35
+ "moe_embedding_experts": true,
36
+ "num_iterations": -1,
37
+ "target_flops": -1.0,
38
+ "target_param_data_ratio": 20,
39
+ "data_dir": "base_data",
40
+ "device_batch_size": 64,
41
+ "total_batch_size": 524288,
42
+ "use_muon": true,
43
+ "embedding_lr": 0.3,
44
+ "unembedding_lr": 0.004,
45
+ "grad_clip": 1.0,
46
+ "weight_decay": 0.0,
47
+ "matrix_lr": 0.02,
48
+ "adam_beta1": 0.8,
49
+ "adam_beta2": 0.95,
50
+ "warmup_ratio": 0.05,
51
+ "warmdown_ratio": 0.4,
52
+ "final_lr_frac": 0.0,
53
+ "resume_from_step": -1,
54
+ "eval_every": 250,
55
+ "eval_tokens": 10485760,
56
+ "core_metric_every": -1,
57
+ "core_metric_max_per_task": 500,
58
+ "sample_every": 250,
59
+ "save_every": 1000,
60
+ "push_checkpoints_to_hub": true,
61
+ "use_profiler": false,
62
+ "profile_step": 2,
63
+ "profile_micro_step": 0,
64
+ "memory_history_max_entries": 10000,
65
+ "model_tag": "moe_2x512",
66
+ "n_layer": 2,
67
+ "n_head": 4,
68
+ "n_kv_head": 4,
69
+ "n_embd": 512
70
+ },
71
+ "user_config": {
72
+ "run": "moe_2x512",
73
+ "wandb_group": null,
74
+ "device_type": "cuda",
75
+ "config": "moe20m",
76
+ "depth": 2,
77
+ "aspect_ratio": 64,
78
+ "model_dim": 512,
79
+ "head_dim": 128,
80
+ "sequence_len": 1024,
81
+ "vocab_size": 265,
82
+ "rva_blocks": [],
83
+ "recurrent_vocab_sizes": [],
84
+ "kla_blocks": [],
85
+ "mamba_blocks": [],
86
+ "gdn_blocks": [],
87
+ "gla_blocks": [],
88
+ "moe_blocks": [
89
+ -1
90
+ ],
91
+ "d_state": 16,
92
+ "mamba_params": true,
93
+ "kla_kernel": true,
94
+ "mimo_rank": 1,
95
+ "skip_around_kla": true,
96
+ "decoder_mlp": false,
97
+ "use_reparametrisation_trick": true,
98
+ "num_experts": 4,
99
+ "moe_top_k": 2,
100
+ "moe_capacity_factor": 1.0,
101
+ "moe_aux_loss_weight": 0.01,
102
+ "moe_embedding_experts": true,
103
+ "num_iterations": -1,
104
+ "target_flops": -1.0,
105
+ "target_param_data_ratio": 20,
106
+ "data_dir": "base_data",
107
+ "device_batch_size": 64,
108
+ "total_batch_size": 524288,
109
+ "use_muon": true,
110
+ "embedding_lr": 0.3,
111
+ "unembedding_lr": 0.004,
112
+ "grad_clip": 1.0,
113
+ "weight_decay": 0.0,
114
+ "matrix_lr": 0.02,
115
+ "adam_beta1": 0.8,
116
+ "adam_beta2": 0.95,
117
+ "warmup_ratio": 0.05,
118
+ "warmdown_ratio": 0.4,
119
+ "final_lr_frac": 0.0,
120
+ "resume_from_step": -1,
121
+ "eval_every": 250,
122
+ "eval_tokens": 10485760,
123
+ "core_metric_every": -1,
124
+ "core_metric_max_per_task": 500,
125
+ "sample_every": 250,
126
+ "save_every": 1000,
127
+ "push_checkpoints_to_hub": true,
128
+ "use_profiler": false,
129
+ "profile_step": 2,
130
+ "profile_micro_step": 0,
131
+ "memory_history_max_entries": 10000,
132
+ "model_tag": "moe_2x512"
133
+ },
134
+ "device_batch_size": 64,
135
+ "sequence_len": 1024,
136
+ "dataloader_state_dict": {
137
+ "pq_idx": 0,
138
+ "rg_idx": 8
139
+ },
140
+ "loop_state": {
141
+ "min_val_bpb": 2.9215333618334247,
142
+ "smooth_train_loss": 2.1010311600203524,
143
+ "smooth_aux_loss": 0.019995352082017165,
144
+ "total_training_time": 7.5558552742004395
145
+ }
146
+ }