Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
d215a73
·
verified ·
1 Parent(s): 3ea2e75

Upload rva_4x256_dhhikf9q/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_4x256_dhhikf9q/meta_002870.json +219 -0
rva_4x256_dhhikf9q/meta_002870.json ADDED
@@ -0,0 +1,219 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": 1.6615904346867678,
4
+ "model_config": {
5
+ "run": "rva_4x256",
6
+ "wandb_group": null,
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": "rvamix_1l",
10
+ "depth": 4,
11
+ "aspect_ratio": 64,
12
+ "model_dim": 256,
13
+ "head_dim": 16,
14
+ "sequence_len": 1024,
15
+ "prefill_len": 0,
16
+ "vocab_size": 265,
17
+ "rva_blocks": [
18
+ 0,
19
+ 1
20
+ ],
21
+ "mamba_blocks": [],
22
+ "gdn_blocks": [],
23
+ "gla_blocks": [],
24
+ "moe_blocks": [],
25
+ "kla_blocks": [],
26
+ "recurrent_vocab_sizes": [
27
+ 265,
28
+ 1024
29
+ ],
30
+ "rva_balance_loss_weight": 0.0,
31
+ "rva_balance_loss_temperature": 1.0,
32
+ "rva_entropy_loss_weight": 0.0,
33
+ "rva_use_ema": true,
34
+ "rva_ema_decay": 0.99,
35
+ "rva_ema_epsilon": 1e-05,
36
+ "rva_use_random_reset": false,
37
+ "rva_reset_threshold": 1.0,
38
+ "rva_reset_prob": 1.0,
39
+ "rva_use_l2_norm": false,
40
+ "rva_use_fsq_ae": false,
41
+ "rva_fsq_levels": [
42
+ 4,
43
+ 4,
44
+ 4,
45
+ 4,
46
+ 4,
47
+ 4,
48
+ 4
49
+ ],
50
+ "rva_recon_loss_init": 1.0,
51
+ "rva_recon_loss_final": 0.01,
52
+ "rva_recon_anneal_fraction": 0.8,
53
+ "rva_n_embd": 0,
54
+ "rva_no_mlp": false,
55
+ "rva_unified_embed": false,
56
+ "rva_factored": true,
57
+ "rva_sub_vocab_size": 0,
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": 2870,
75
+ "target_flops": -1.0,
76
+ "target_param_data_ratio": 20,
77
+ "data_dir": "base_data",
78
+ "device_batch_size": 64,
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_4x256",
101
+ "stage": null,
102
+ "n_layer": 4,
103
+ "n_head": 16,
104
+ "n_kv_head": 16,
105
+ "n_embd": 256
106
+ },
107
+ "user_config": {
108
+ "run": "rva_4x256",
109
+ "wandb_group": null,
110
+ "seed": 42,
111
+ "device_type": "cuda",
112
+ "config": "rvamix_1l",
113
+ "depth": 4,
114
+ "aspect_ratio": 64,
115
+ "model_dim": 256,
116
+ "head_dim": 16,
117
+ "sequence_len": 1024,
118
+ "prefill_len": 0,
119
+ "vocab_size": 265,
120
+ "rva_blocks": [
121
+ 0,
122
+ 1
123
+ ],
124
+ "mamba_blocks": [],
125
+ "gdn_blocks": [],
126
+ "gla_blocks": [],
127
+ "moe_blocks": [],
128
+ "kla_blocks": [],
129
+ "recurrent_vocab_sizes": [
130
+ 265,
131
+ 1024
132
+ ],
133
+ "rva_balance_loss_weight": 0.0,
134
+ "rva_balance_loss_temperature": 1.0,
135
+ "rva_entropy_loss_weight": 0.0,
136
+ "rva_use_ema": true,
137
+ "rva_ema_decay": 0.99,
138
+ "rva_ema_epsilon": 1e-05,
139
+ "rva_use_random_reset": false,
140
+ "rva_reset_threshold": 1.0,
141
+ "rva_reset_prob": 1.0,
142
+ "rva_use_l2_norm": false,
143
+ "rva_use_fsq_ae": false,
144
+ "rva_fsq_levels": [
145
+ 4,
146
+ 4,
147
+ 4,
148
+ 4,
149
+ 4,
150
+ 4,
151
+ 4
152
+ ],
153
+ "rva_recon_loss_init": 1.0,
154
+ "rva_recon_loss_final": 0.01,
155
+ "rva_recon_anneal_fraction": 0.8,
156
+ "rva_n_embd": 0,
157
+ "rva_no_mlp": false,
158
+ "rva_unified_embed": false,
159
+ "rva_factored": true,
160
+ "rva_sub_vocab_size": 0,
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": 2870,
178
+ "target_flops": -1.0,
179
+ "target_param_data_ratio": 20,
180
+ "data_dir": "base_data",
181
+ "device_batch_size": 64,
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_4x256",
204
+ "stage": null
205
+ },
206
+ "stage": null,
207
+ "device_batch_size": 64,
208
+ "sequence_len": 1024,
209
+ "dataloader_state_dict": {
210
+ "pq_idx": 5,
211
+ "rg_idx": 32
212
+ },
213
+ "loop_state": {
214
+ "min_val_bpb": 1.6615904346867678,
215
+ "smooth_train_loss": 1.170383956381217,
216
+ "smooth_aux_loss": 0.029060003523951338,
217
+ "total_training_time": 351.0196623802185
218
+ }
219
+ }