Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
Community
ASzecsenyi commited on
Commit
9582271
·
verified ·
1 Parent(s): a5416c8

Upload rva_6x256_gh9pvuzz/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_6x256_gh9pvuzz/meta_002870.json +218 -0
rva_6x256_gh9pvuzz/meta_002870.json ADDED
@@ -0,0 +1,218 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": NaN,
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
+ ],
22
+ "mamba_blocks": [],
23
+ "gdn_blocks": [],
24
+ "gla_blocks": [],
25
+ "moe_blocks": [],
26
+ "kla_blocks": [],
27
+ "recurrent_vocab_sizes": [
28
+ 16383,
29
+ 16384,
30
+ 16384,
31
+ 16384
32
+ ],
33
+ "rva_balance_loss_weight": 0.0,
34
+ "rva_balance_loss_temperature": 1.0,
35
+ "rva_entropy_loss_weight": 0.0,
36
+ "rva_use_ema": true,
37
+ "rva_ema_decay": 0.99,
38
+ "rva_ema_epsilon": 1e-05,
39
+ "rva_use_random_reset": false,
40
+ "rva_reset_threshold": 1.0,
41
+ "rva_reset_prob": 1.0,
42
+ "rva_use_l2_norm": false,
43
+ "rva_use_fft_conv": true,
44
+ "rva_use_fsq_ae": false,
45
+ "rva_fsq_levels": [
46
+ 4,
47
+ 4,
48
+ 4,
49
+ 4,
50
+ 4,
51
+ 4,
52
+ 4
53
+ ],
54
+ "rva_recon_loss_init": 1.0,
55
+ "rva_recon_loss_final": 0.01,
56
+ "rva_recon_anneal_fraction": 0.8,
57
+ "use_skip_connection": true,
58
+ "clustering_init_num_samples": 0,
59
+ "max_active_tokens_in_batch": -1,
60
+ "max_active_token_fraction": 0.0,
61
+ "d_state": 16,
62
+ "mamba_params": true,
63
+ "kla_kernel": true,
64
+ "mimo_rank": 1,
65
+ "skip_around_kla": true,
66
+ "decoder_mlp": false,
67
+ "use_reparametrisation_trick": true,
68
+ "num_experts": 4,
69
+ "moe_top_k": 2,
70
+ "moe_capacity_factor": 1.0,
71
+ "moe_aux_loss_weight": 0.01,
72
+ "moe_embedding_experts": false,
73
+ "num_iterations": -1,
74
+ "target_flops": -1.0,
75
+ "target_param_data_ratio": 20,
76
+ "data_dir": "base_data",
77
+ "device_batch_size": 8,
78
+ "total_batch_size": 524288,
79
+ "use_muon": true,
80
+ "embedding_lr": 0.3,
81
+ "unembedding_lr": 0.004,
82
+ "grad_clip": 1.0,
83
+ "weight_decay": 0.0,
84
+ "matrix_lr": 0.02,
85
+ "adam_beta1": 0.8,
86
+ "adam_beta2": 0.95,
87
+ "warmup_ratio": 0.05,
88
+ "warmdown_ratio": 0.4,
89
+ "final_lr_frac": 0.0,
90
+ "resume_from_step": -1,
91
+ "eval_every": 250,
92
+ "eval_tokens": 10485760,
93
+ "core_metric_every": -1,
94
+ "core_metric_max_per_task": 500,
95
+ "sample_every": 250,
96
+ "save_every": 1000,
97
+ "push_checkpoints_to_hub": true,
98
+ "profile_rva_codebook_usage": true,
99
+ "model_tag": "rva_6x256",
100
+ "stage": null,
101
+ "n_layer": 6,
102
+ "n_head": 16,
103
+ "n_kv_head": 16,
104
+ "n_embd": 256
105
+ },
106
+ "user_config": {
107
+ "run": "rva_6x256",
108
+ "wandb_group": null,
109
+ "seed": 42,
110
+ "device_type": "cuda",
111
+ "config": "rvamix5m_v16k",
112
+ "depth": 6,
113
+ "aspect_ratio": 64,
114
+ "model_dim": 256,
115
+ "head_dim": 16,
116
+ "sequence_len": 1024,
117
+ "vocab_size": 16384,
118
+ "rva_blocks": [
119
+ 0,
120
+ 1,
121
+ 2,
122
+ 3
123
+ ],
124
+ "mamba_blocks": [],
125
+ "gdn_blocks": [],
126
+ "gla_blocks": [],
127
+ "moe_blocks": [],
128
+ "kla_blocks": [],
129
+ "recurrent_vocab_sizes": [
130
+ 16383,
131
+ 16384,
132
+ 16384,
133
+ 16384
134
+ ],
135
+ "rva_balance_loss_weight": 0.0,
136
+ "rva_balance_loss_temperature": 1.0,
137
+ "rva_entropy_loss_weight": 0.0,
138
+ "rva_use_ema": true,
139
+ "rva_ema_decay": 0.99,
140
+ "rva_ema_epsilon": 1e-05,
141
+ "rva_use_random_reset": false,
142
+ "rva_reset_threshold": 1.0,
143
+ "rva_reset_prob": 1.0,
144
+ "rva_use_l2_norm": false,
145
+ "rva_use_fft_conv": true,
146
+ "rva_use_fsq_ae": false,
147
+ "rva_fsq_levels": [
148
+ 4,
149
+ 4,
150
+ 4,
151
+ 4,
152
+ 4,
153
+ 4,
154
+ 4
155
+ ],
156
+ "rva_recon_loss_init": 1.0,
157
+ "rva_recon_loss_final": 0.01,
158
+ "rva_recon_anneal_fraction": 0.8,
159
+ "use_skip_connection": true,
160
+ "clustering_init_num_samples": 0,
161
+ "max_active_tokens_in_batch": -1,
162
+ "max_active_token_fraction": 0.0,
163
+ "d_state": 16,
164
+ "mamba_params": true,
165
+ "kla_kernel": true,
166
+ "mimo_rank": 1,
167
+ "skip_around_kla": true,
168
+ "decoder_mlp": false,
169
+ "use_reparametrisation_trick": true,
170
+ "num_experts": 4,
171
+ "moe_top_k": 2,
172
+ "moe_capacity_factor": 1.0,
173
+ "moe_aux_loss_weight": 0.01,
174
+ "moe_embedding_experts": false,
175
+ "num_iterations": -1,
176
+ "target_flops": -1.0,
177
+ "target_param_data_ratio": 20,
178
+ "data_dir": "base_data",
179
+ "device_batch_size": 8,
180
+ "total_batch_size": 524288,
181
+ "use_muon": true,
182
+ "embedding_lr": 0.3,
183
+ "unembedding_lr": 0.004,
184
+ "grad_clip": 1.0,
185
+ "weight_decay": 0.0,
186
+ "matrix_lr": 0.02,
187
+ "adam_beta1": 0.8,
188
+ "adam_beta2": 0.95,
189
+ "warmup_ratio": 0.05,
190
+ "warmdown_ratio": 0.4,
191
+ "final_lr_frac": 0.0,
192
+ "resume_from_step": -1,
193
+ "eval_every": 250,
194
+ "eval_tokens": 10485760,
195
+ "core_metric_every": -1,
196
+ "core_metric_max_per_task": 500,
197
+ "sample_every": 250,
198
+ "save_every": 1000,
199
+ "push_checkpoints_to_hub": true,
200
+ "profile_rva_codebook_usage": true,
201
+ "model_tag": "rva_6x256",
202
+ "stage": null,
203
+ "pod_name": "s2027538-infk8s-job-r4d2k-mvjw4"
204
+ },
205
+ "stage": null,
206
+ "device_batch_size": 8,
207
+ "sequence_len": 1024,
208
+ "dataloader_state_dict": {
209
+ "pq_idx": 25,
210
+ "rg_idx": 30
211
+ },
212
+ "loop_state": {
213
+ "min_val_bpb": 1.5163627796975052,
214
+ "smooth_train_loss": NaN,
215
+ "smooth_aux_loss": NaN,
216
+ "total_training_time": 18931.141661643982
217
+ }
218
+ }