Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
2970300
·
verified ·
1 Parent(s): cbeaaf6

Upload rva_6x256_q72tnbsq/meta_003502.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_6x256_q72tnbsq/meta_003502.json +220 -0
rva_6x256_q72tnbsq/meta_003502.json ADDED
@@ -0,0 +1,220 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 3502,
3
+ "val_bpb": 1.4914860930340665,
4
+ "model_config": {
5
+ "run": "rva_6x256",
6
+ "wandb_group": null,
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": "rvamix5m_v16k",
10
+ "depth": 6,
11
+ "aspect_ratio": 64,
12
+ "model_dim": 256,
13
+ "head_dim": 16,
14
+ "sequence_len": 1024,
15
+ "vocab_size": 16384,
16
+ "rva_blocks": [
17
+ 0,
18
+ 1,
19
+ 2,
20
+ 3,
21
+ 4
22
+ ],
23
+ "mamba_blocks": [],
24
+ "gdn_blocks": [],
25
+ "gla_blocks": [],
26
+ "moe_blocks": [],
27
+ "kla_blocks": [],
28
+ "recurrent_vocab_sizes": [
29
+ 16383,
30
+ 16384,
31
+ 16384,
32
+ 16384,
33
+ 16384
34
+ ],
35
+ "rva_balance_loss_weight": 0.0,
36
+ "rva_balance_loss_temperature": 1.0,
37
+ "rva_entropy_loss_weight": 0.0,
38
+ "rva_use_ema": true,
39
+ "rva_ema_decay": 0.99,
40
+ "rva_ema_epsilon": 1e-05,
41
+ "rva_use_random_reset": false,
42
+ "rva_reset_threshold": 1.0,
43
+ "rva_reset_prob": 1.0,
44
+ "rva_use_l2_norm": false,
45
+ "rva_use_fsq_ae": false,
46
+ "rva_fsq_levels": [
47
+ 4,
48
+ 4,
49
+ 4,
50
+ 4,
51
+ 4,
52
+ 4,
53
+ 4
54
+ ],
55
+ "rva_recon_loss_init": 1.0,
56
+ "rva_recon_loss_final": 0.01,
57
+ "rva_recon_anneal_fraction": 0.8,
58
+ "use_skip_connection": true,
59
+ "clustering_init_num_samples": 0,
60
+ "max_active_tokens_in_batch": -1,
61
+ "max_active_token_fraction": 0.0,
62
+ "d_state": 16,
63
+ "mamba_params": true,
64
+ "kla_kernel": true,
65
+ "mimo_rank": 1,
66
+ "skip_around_kla": true,
67
+ "decoder_mlp": false,
68
+ "use_reparametrisation_trick": true,
69
+ "num_experts": 4,
70
+ "moe_top_k": 2,
71
+ "moe_capacity_factor": 1.0,
72
+ "moe_aux_loss_weight": 0.01,
73
+ "moe_embedding_experts": false,
74
+ "num_iterations": -1,
75
+ "target_flops": -1.0,
76
+ "target_param_data_ratio": 20,
77
+ "data_dir": "base_data",
78
+ "device_batch_size": 8,
79
+ "total_batch_size": 524288,
80
+ "use_muon": true,
81
+ "embedding_lr": 0.3,
82
+ "unembedding_lr": 0.004,
83
+ "grad_clip": 1.0,
84
+ "weight_decay": 0.0,
85
+ "matrix_lr": 0.02,
86
+ "adam_beta1": 0.8,
87
+ "adam_beta2": 0.95,
88
+ "warmup_ratio": 0.05,
89
+ "warmdown_ratio": 0.4,
90
+ "final_lr_frac": 0.0,
91
+ "resume_from_step": -1,
92
+ "eval_every": 250,
93
+ "eval_tokens": 10485760,
94
+ "core_metric_every": -1,
95
+ "core_metric_max_per_task": 500,
96
+ "sample_every": 250,
97
+ "save_every": 1000,
98
+ "push_checkpoints_to_hub": true,
99
+ "profile_rva_codebook_usage": true,
100
+ "model_tag": "rva_6x256",
101
+ "stage": null,
102
+ "n_layer": 6,
103
+ "n_head": 16,
104
+ "n_kv_head": 16,
105
+ "n_embd": 256
106
+ },
107
+ "user_config": {
108
+ "run": "rva_6x256",
109
+ "wandb_group": null,
110
+ "seed": 42,
111
+ "device_type": "cuda",
112
+ "config": "rvamix5m_v16k",
113
+ "depth": 6,
114
+ "aspect_ratio": 64,
115
+ "model_dim": 256,
116
+ "head_dim": 16,
117
+ "sequence_len": 1024,
118
+ "vocab_size": 16384,
119
+ "rva_blocks": [
120
+ 0,
121
+ 1,
122
+ 2,
123
+ 3,
124
+ 4
125
+ ],
126
+ "mamba_blocks": [],
127
+ "gdn_blocks": [],
128
+ "gla_blocks": [],
129
+ "moe_blocks": [],
130
+ "kla_blocks": [],
131
+ "recurrent_vocab_sizes": [
132
+ 16383,
133
+ 16384,
134
+ 16384,
135
+ 16384,
136
+ 16384
137
+ ],
138
+ "rva_balance_loss_weight": 0.0,
139
+ "rva_balance_loss_temperature": 1.0,
140
+ "rva_entropy_loss_weight": 0.0,
141
+ "rva_use_ema": true,
142
+ "rva_ema_decay": 0.99,
143
+ "rva_ema_epsilon": 1e-05,
144
+ "rva_use_random_reset": false,
145
+ "rva_reset_threshold": 1.0,
146
+ "rva_reset_prob": 1.0,
147
+ "rva_use_l2_norm": false,
148
+ "rva_use_fsq_ae": false,
149
+ "rva_fsq_levels": [
150
+ 4,
151
+ 4,
152
+ 4,
153
+ 4,
154
+ 4,
155
+ 4,
156
+ 4
157
+ ],
158
+ "rva_recon_loss_init": 1.0,
159
+ "rva_recon_loss_final": 0.01,
160
+ "rva_recon_anneal_fraction": 0.8,
161
+ "use_skip_connection": true,
162
+ "clustering_init_num_samples": 0,
163
+ "max_active_tokens_in_batch": -1,
164
+ "max_active_token_fraction": 0.0,
165
+ "d_state": 16,
166
+ "mamba_params": true,
167
+ "kla_kernel": true,
168
+ "mimo_rank": 1,
169
+ "skip_around_kla": true,
170
+ "decoder_mlp": false,
171
+ "use_reparametrisation_trick": true,
172
+ "num_experts": 4,
173
+ "moe_top_k": 2,
174
+ "moe_capacity_factor": 1.0,
175
+ "moe_aux_loss_weight": 0.01,
176
+ "moe_embedding_experts": false,
177
+ "num_iterations": -1,
178
+ "target_flops": -1.0,
179
+ "target_param_data_ratio": 20,
180
+ "data_dir": "base_data",
181
+ "device_batch_size": 8,
182
+ "total_batch_size": 524288,
183
+ "use_muon": true,
184
+ "embedding_lr": 0.3,
185
+ "unembedding_lr": 0.004,
186
+ "grad_clip": 1.0,
187
+ "weight_decay": 0.0,
188
+ "matrix_lr": 0.02,
189
+ "adam_beta1": 0.8,
190
+ "adam_beta2": 0.95,
191
+ "warmup_ratio": 0.05,
192
+ "warmdown_ratio": 0.4,
193
+ "final_lr_frac": 0.0,
194
+ "resume_from_step": -1,
195
+ "eval_every": 250,
196
+ "eval_tokens": 10485760,
197
+ "core_metric_every": -1,
198
+ "core_metric_max_per_task": 500,
199
+ "sample_every": 250,
200
+ "save_every": 1000,
201
+ "push_checkpoints_to_hub": true,
202
+ "profile_rva_codebook_usage": true,
203
+ "model_tag": "rva_6x256",
204
+ "stage": null,
205
+ "pod_name": "s2027538-infk8s-job-966bg-lz7dn"
206
+ },
207
+ "stage": null,
208
+ "device_batch_size": 8,
209
+ "sequence_len": 1024,
210
+ "dataloader_state_dict": {
211
+ "pq_idx": 31,
212
+ "rg_idx": 12
213
+ },
214
+ "loop_state": {
215
+ "min_val_bpb": 1.4914860930340665,
216
+ "smooth_train_loss": 4.964072982337414,
217
+ "smooth_aux_loss": 0.3684133280917994,
218
+ "total_training_time": 28488.441681861877
219
+ }
220
+ }