Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
da9379e
·
verified ·
1 Parent(s): 9f8518c

Upload 1rva_2gpt_v1024_noinit_57ermfqd/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. 1rva_2gpt_v1024_noinit_57ermfqd/meta_002870.json +235 -0
1rva_2gpt_v1024_noinit_57ermfqd/meta_002870.json ADDED
@@ -0,0 +1,235 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": 1.3631244788886434,
4
+ "model_config": {
5
+ "run": "learned_init_ablation/1rva_2gpt_v1024_noinit",
6
+ "wandb_group": "learned_init_ablation",
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": null,
10
+ "depth": 3,
11
+ "aspect_ratio": 64,
12
+ "model_dim": 256,
13
+ "head_dim": 64,
14
+ "sequence_len": 1024,
15
+ "prefill_len": 0,
16
+ "vocab_size": 1024,
17
+ "rva_blocks": [
18
+ 0
19
+ ],
20
+ "mamba_blocks": [],
21
+ "gdn_blocks": [],
22
+ "gla_blocks": [],
23
+ "moe_blocks": [],
24
+ "kla_blocks": [],
25
+ "recurrent_vocab_sizes": [
26
+ 1024
27
+ ],
28
+ "rva_balance_loss_weight": 0.0,
29
+ "rva_balance_loss_temperature": 1.0,
30
+ "rva_entropy_loss_weight": 0.0,
31
+ "rva_use_ema": true,
32
+ "rva_ema_decay": 0.99,
33
+ "rva_ema_epsilon": 1e-05,
34
+ "rva_use_random_reset": false,
35
+ "rva_reset_threshold": 1.0,
36
+ "rva_reset_prob": 1.0,
37
+ "rva_use_l2_norm": false,
38
+ "rva_use_fsq_ae": false,
39
+ "rva_fsq_levels": [
40
+ 4,
41
+ 4,
42
+ 4,
43
+ 4,
44
+ 4,
45
+ 4,
46
+ 4
47
+ ],
48
+ "rva_recon_loss_init": 1.0,
49
+ "rva_recon_loss_final": 0.01,
50
+ "rva_recon_anneal_fraction": 0.8,
51
+ "rva_n_embd": 0,
52
+ "rva_no_mlp": false,
53
+ "rva_unified_embed": false,
54
+ "rva_embed_in_muon": true,
55
+ "rva_learned_init_state": false,
56
+ "rva_pq_groups": 0,
57
+ "rva_pq_k_sub": 0,
58
+ "rva_pq_balance_loss_weight": 0.0,
59
+ "rva_use_gumbel": false,
60
+ "rva_gumbel_temp": 2.0,
61
+ "rva_gumbel_temp_min": 0.5,
62
+ "rva_gumbel_anneal_steps": 0,
63
+ "rva_gumbel_diversity_weight": 0.1,
64
+ "rva_factored": false,
65
+ "rva_sub_vocab_size": 0,
66
+ "use_skip_connection": true,
67
+ "clustering_init_num_samples": 10000,
68
+ "max_active_tokens_in_batch": -1,
69
+ "max_active_token_fraction": 0.0,
70
+ "d_state": 16,
71
+ "mamba_params": true,
72
+ "kla_kernel": true,
73
+ "mimo_rank": 1,
74
+ "skip_around_kla": true,
75
+ "decoder_mlp": false,
76
+ "use_reparametrisation_trick": true,
77
+ "num_experts": 4,
78
+ "moe_top_k": 2,
79
+ "moe_capacity_factor": 1.0,
80
+ "moe_aux_loss_weight": 0.01,
81
+ "moe_embedding_experts": false,
82
+ "num_iterations": 2870,
83
+ "target_flops": -1.0,
84
+ "target_param_data_ratio": 20,
85
+ "data_dir": "base_data",
86
+ "device_batch_size": 64,
87
+ "total_batch_size": 524288,
88
+ "use_muon": true,
89
+ "embedding_lr": 0.3,
90
+ "unembedding_lr": 0.004,
91
+ "grad_clip": 1.0,
92
+ "weight_decay": 0.0,
93
+ "matrix_lr": 0.02,
94
+ "adam_beta1": 0.8,
95
+ "adam_beta2": 0.95,
96
+ "warmup_ratio": 0.05,
97
+ "warmdown_ratio": 0.4,
98
+ "final_lr_frac": 0.0,
99
+ "resume_from_step": -1,
100
+ "eval_every": 250,
101
+ "eval_tokens": 10485760,
102
+ "core_metric_every": -1,
103
+ "core_metric_max_per_task": 500,
104
+ "sample_every": 250,
105
+ "save_every": 1000,
106
+ "push_checkpoints_to_hub": true,
107
+ "profile_rva_codebook_usage": true,
108
+ "model_tag": "learned_init_ablation/1rva_2gpt_v1024_noinit",
109
+ "stage": null,
110
+ "n_layer": 3,
111
+ "n_head": 4,
112
+ "n_kv_head": 4,
113
+ "n_embd": 256
114
+ },
115
+ "user_config": {
116
+ "run": "learned_init_ablation/1rva_2gpt_v1024_noinit",
117
+ "wandb_group": "learned_init_ablation",
118
+ "seed": 42,
119
+ "device_type": "cuda",
120
+ "config": null,
121
+ "depth": 3,
122
+ "aspect_ratio": 64,
123
+ "model_dim": 256,
124
+ "head_dim": 64,
125
+ "sequence_len": 1024,
126
+ "prefill_len": 0,
127
+ "vocab_size": 1024,
128
+ "rva_blocks": [
129
+ 0
130
+ ],
131
+ "mamba_blocks": [],
132
+ "gdn_blocks": [],
133
+ "gla_blocks": [],
134
+ "moe_blocks": [],
135
+ "kla_blocks": [],
136
+ "recurrent_vocab_sizes": [
137
+ 1024
138
+ ],
139
+ "rva_balance_loss_weight": 0.0,
140
+ "rva_balance_loss_temperature": 1.0,
141
+ "rva_entropy_loss_weight": 0.0,
142
+ "rva_use_ema": true,
143
+ "rva_ema_decay": 0.99,
144
+ "rva_ema_epsilon": 1e-05,
145
+ "rva_use_random_reset": false,
146
+ "rva_reset_threshold": 1.0,
147
+ "rva_reset_prob": 1.0,
148
+ "rva_use_l2_norm": false,
149
+ "rva_use_fsq_ae": false,
150
+ "rva_fsq_levels": [
151
+ 4,
152
+ 4,
153
+ 4,
154
+ 4,
155
+ 4,
156
+ 4,
157
+ 4
158
+ ],
159
+ "rva_recon_loss_init": 1.0,
160
+ "rva_recon_loss_final": 0.01,
161
+ "rva_recon_anneal_fraction": 0.8,
162
+ "rva_n_embd": 0,
163
+ "rva_no_mlp": false,
164
+ "rva_unified_embed": false,
165
+ "rva_embed_in_muon": true,
166
+ "rva_learned_init_state": false,
167
+ "rva_pq_groups": 0,
168
+ "rva_pq_k_sub": 0,
169
+ "rva_pq_balance_loss_weight": 0.0,
170
+ "rva_use_gumbel": false,
171
+ "rva_gumbel_temp": 2.0,
172
+ "rva_gumbel_temp_min": 0.5,
173
+ "rva_gumbel_anneal_steps": 0,
174
+ "rva_gumbel_diversity_weight": 0.1,
175
+ "rva_factored": false,
176
+ "rva_sub_vocab_size": 0,
177
+ "use_skip_connection": true,
178
+ "clustering_init_num_samples": 10000,
179
+ "max_active_tokens_in_batch": -1,
180
+ "max_active_token_fraction": 0.0,
181
+ "d_state": 16,
182
+ "mamba_params": true,
183
+ "kla_kernel": true,
184
+ "mimo_rank": 1,
185
+ "skip_around_kla": true,
186
+ "decoder_mlp": false,
187
+ "use_reparametrisation_trick": true,
188
+ "num_experts": 4,
189
+ "moe_top_k": 2,
190
+ "moe_capacity_factor": 1.0,
191
+ "moe_aux_loss_weight": 0.01,
192
+ "moe_embedding_experts": false,
193
+ "num_iterations": 2870,
194
+ "target_flops": -1.0,
195
+ "target_param_data_ratio": 20,
196
+ "data_dir": "base_data",
197
+ "device_batch_size": 64,
198
+ "total_batch_size": 524288,
199
+ "use_muon": true,
200
+ "embedding_lr": 0.3,
201
+ "unembedding_lr": 0.004,
202
+ "grad_clip": 1.0,
203
+ "weight_decay": 0.0,
204
+ "matrix_lr": 0.02,
205
+ "adam_beta1": 0.8,
206
+ "adam_beta2": 0.95,
207
+ "warmup_ratio": 0.05,
208
+ "warmdown_ratio": 0.4,
209
+ "final_lr_frac": 0.0,
210
+ "resume_from_step": -1,
211
+ "eval_every": 250,
212
+ "eval_tokens": 10485760,
213
+ "core_metric_every": -1,
214
+ "core_metric_max_per_task": 500,
215
+ "sample_every": 250,
216
+ "save_every": 1000,
217
+ "push_checkpoints_to_hub": true,
218
+ "profile_rva_codebook_usage": true,
219
+ "model_tag": "learned_init_ablation/1rva_2gpt_v1024_noinit",
220
+ "stage": null
221
+ },
222
+ "stage": null,
223
+ "device_batch_size": 64,
224
+ "sequence_len": 1024,
225
+ "dataloader_state_dict": {
226
+ "pq_idx": 13,
227
+ "rg_idx": 48
228
+ },
229
+ "loop_state": {
230
+ "min_val_bpb": 1.3631244788886434,
231
+ "smooth_train_loss": 2.406446279649803,
232
+ "smooth_aux_loss": 0.0,
233
+ "total_training_time": 331.44384360313416
234
+ }
235
+ }