Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
0d34336
·
verified ·
1 Parent(s): 49a8b14

Upload 2rva_2gpt_gs_8kcodes_t128mdn0/meta_001000.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. 2rva_2gpt_gs_8kcodes_t128mdn0/meta_001000.json +241 -0
2rva_2gpt_gs_8kcodes_t128mdn0/meta_001000.json ADDED
@@ -0,0 +1,241 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 1000,
3
+ "val_bpb": 4.591894460576155,
4
+ "model_config": {
5
+ "run": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
6
+ "wandb_group": "gumbel_ablation",
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": null,
10
+ "depth": 5,
11
+ "aspect_ratio": 64,
12
+ "model_dim": 256,
13
+ "head_dim": 64,
14
+ "sequence_len": 1024,
15
+ "prefill_len": 0,
16
+ "vocab_size": 265,
17
+ "rva_blocks": [
18
+ 0,
19
+ 1,
20
+ 2
21
+ ],
22
+ "mamba_blocks": [],
23
+ "gdn_blocks": [],
24
+ "gla_blocks": [],
25
+ "moe_blocks": [],
26
+ "kla_blocks": [],
27
+ "recurrent_vocab_sizes": [
28
+ 265,
29
+ 8192,
30
+ 8192
31
+ ],
32
+ "rva_balance_loss_weight": 0.0,
33
+ "rva_balance_loss_temperature": 1.0,
34
+ "rva_entropy_loss_weight": 0.0,
35
+ "rva_use_ema": true,
36
+ "rva_ema_decay": 0.99,
37
+ "rva_ema_epsilon": 1e-05,
38
+ "rva_use_random_reset": false,
39
+ "rva_reset_threshold": 1.0,
40
+ "rva_reset_prob": 1.0,
41
+ "rva_use_l2_norm": false,
42
+ "rva_use_fsq_ae": false,
43
+ "rva_fsq_levels": [
44
+ 4,
45
+ 4,
46
+ 4,
47
+ 4,
48
+ 4,
49
+ 4,
50
+ 4
51
+ ],
52
+ "rva_recon_loss_init": 1.0,
53
+ "rva_recon_loss_final": 0.01,
54
+ "rva_recon_anneal_fraction": 0.8,
55
+ "rva_n_embd": 0,
56
+ "rva_no_mlp": false,
57
+ "rva_unified_embed": false,
58
+ "rva_embed_in_muon": true,
59
+ "rva_pq_groups": 0,
60
+ "rva_pq_k_sub": 0,
61
+ "rva_pq_balance_loss_weight": 0.0,
62
+ "rva_use_gumbel": true,
63
+ "rva_gumbel_temp": 2.0,
64
+ "rva_gumbel_temp_min": 0.1,
65
+ "rva_gumbel_anneal_steps": 0,
66
+ "rva_gumbel_diversity_weight": 0.001,
67
+ "rva_factored": false,
68
+ "rva_sub_vocab_size": 0,
69
+ "use_skip_connection": true,
70
+ "clustering_init_num_samples": 0,
71
+ "max_active_tokens_in_batch": -1,
72
+ "max_active_token_fraction": 0.0,
73
+ "d_state": 16,
74
+ "mamba_params": true,
75
+ "kla_kernel": true,
76
+ "mimo_rank": 1,
77
+ "skip_around_kla": true,
78
+ "decoder_mlp": false,
79
+ "use_reparametrisation_trick": true,
80
+ "num_experts": 4,
81
+ "moe_top_k": 2,
82
+ "moe_capacity_factor": 1.0,
83
+ "moe_aux_loss_weight": 0.01,
84
+ "moe_embedding_experts": false,
85
+ "num_iterations": 2870,
86
+ "target_flops": -1.0,
87
+ "target_param_data_ratio": 20,
88
+ "data_dir": "base_data",
89
+ "device_batch_size": 8,
90
+ "total_batch_size": 524288,
91
+ "use_muon": true,
92
+ "embedding_lr": 0.3,
93
+ "unembedding_lr": 0.004,
94
+ "grad_clip": 1.0,
95
+ "weight_decay": 0.0,
96
+ "matrix_lr": 0.02,
97
+ "adam_beta1": 0.8,
98
+ "adam_beta2": 0.95,
99
+ "warmup_ratio": 0.05,
100
+ "warmdown_ratio": 0.4,
101
+ "final_lr_frac": 0.0,
102
+ "resume_from_step": -1,
103
+ "eval_every": 250,
104
+ "eval_tokens": 10485760,
105
+ "core_metric_every": -1,
106
+ "core_metric_max_per_task": 500,
107
+ "sample_every": 250,
108
+ "save_every": 1000,
109
+ "push_checkpoints_to_hub": true,
110
+ "profile_rva_codebook_usage": true,
111
+ "model_tag": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
112
+ "stage": null,
113
+ "n_layer": 5,
114
+ "n_head": 4,
115
+ "n_kv_head": 4,
116
+ "n_embd": 256
117
+ },
118
+ "user_config": {
119
+ "run": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
120
+ "wandb_group": "gumbel_ablation",
121
+ "seed": 42,
122
+ "device_type": "cuda",
123
+ "config": null,
124
+ "depth": 5,
125
+ "aspect_ratio": 64,
126
+ "model_dim": 256,
127
+ "head_dim": 64,
128
+ "sequence_len": 1024,
129
+ "prefill_len": 0,
130
+ "vocab_size": 265,
131
+ "rva_blocks": [
132
+ 0,
133
+ 1,
134
+ 2
135
+ ],
136
+ "mamba_blocks": [],
137
+ "gdn_blocks": [],
138
+ "gla_blocks": [],
139
+ "moe_blocks": [],
140
+ "kla_blocks": [],
141
+ "recurrent_vocab_sizes": [
142
+ 265,
143
+ 8192,
144
+ 8192
145
+ ],
146
+ "rva_balance_loss_weight": 0.0,
147
+ "rva_balance_loss_temperature": 1.0,
148
+ "rva_entropy_loss_weight": 0.0,
149
+ "rva_use_ema": true,
150
+ "rva_ema_decay": 0.99,
151
+ "rva_ema_epsilon": 1e-05,
152
+ "rva_use_random_reset": false,
153
+ "rva_reset_threshold": 1.0,
154
+ "rva_reset_prob": 1.0,
155
+ "rva_use_l2_norm": false,
156
+ "rva_use_fsq_ae": false,
157
+ "rva_fsq_levels": [
158
+ 4,
159
+ 4,
160
+ 4,
161
+ 4,
162
+ 4,
163
+ 4,
164
+ 4
165
+ ],
166
+ "rva_recon_loss_init": 1.0,
167
+ "rva_recon_loss_final": 0.01,
168
+ "rva_recon_anneal_fraction": 0.8,
169
+ "rva_n_embd": 0,
170
+ "rva_no_mlp": false,
171
+ "rva_unified_embed": false,
172
+ "rva_embed_in_muon": true,
173
+ "rva_pq_groups": 0,
174
+ "rva_pq_k_sub": 0,
175
+ "rva_pq_balance_loss_weight": 0.0,
176
+ "rva_use_gumbel": true,
177
+ "rva_gumbel_temp": 2.0,
178
+ "rva_gumbel_temp_min": 0.1,
179
+ "rva_gumbel_anneal_steps": 0,
180
+ "rva_gumbel_diversity_weight": 0.001,
181
+ "rva_factored": false,
182
+ "rva_sub_vocab_size": 0,
183
+ "use_skip_connection": true,
184
+ "clustering_init_num_samples": 0,
185
+ "max_active_tokens_in_batch": -1,
186
+ "max_active_token_fraction": 0.0,
187
+ "d_state": 16,
188
+ "mamba_params": true,
189
+ "kla_kernel": true,
190
+ "mimo_rank": 1,
191
+ "skip_around_kla": true,
192
+ "decoder_mlp": false,
193
+ "use_reparametrisation_trick": true,
194
+ "num_experts": 4,
195
+ "moe_top_k": 2,
196
+ "moe_capacity_factor": 1.0,
197
+ "moe_aux_loss_weight": 0.01,
198
+ "moe_embedding_experts": false,
199
+ "num_iterations": 2870,
200
+ "target_flops": -1.0,
201
+ "target_param_data_ratio": 20,
202
+ "data_dir": "base_data",
203
+ "device_batch_size": 8,
204
+ "total_batch_size": 524288,
205
+ "use_muon": true,
206
+ "embedding_lr": 0.3,
207
+ "unembedding_lr": 0.004,
208
+ "grad_clip": 1.0,
209
+ "weight_decay": 0.0,
210
+ "matrix_lr": 0.02,
211
+ "adam_beta1": 0.8,
212
+ "adam_beta2": 0.95,
213
+ "warmup_ratio": 0.05,
214
+ "warmdown_ratio": 0.4,
215
+ "final_lr_frac": 0.0,
216
+ "resume_from_step": -1,
217
+ "eval_every": 250,
218
+ "eval_tokens": 10485760,
219
+ "core_metric_every": -1,
220
+ "core_metric_max_per_task": 500,
221
+ "sample_every": 250,
222
+ "save_every": 1000,
223
+ "push_checkpoints_to_hub": true,
224
+ "profile_rva_codebook_usage": true,
225
+ "model_tag": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
226
+ "stage": null
227
+ },
228
+ "stage": null,
229
+ "device_batch_size": 8,
230
+ "sequence_len": 1024,
231
+ "dataloader_state_dict": {
232
+ "pq_idx": 1,
233
+ "rg_idx": 48
234
+ },
235
+ "loop_state": {
236
+ "min_val_bpb": 2.6370529804971157,
237
+ "smooth_train_loss": 3.0919047934307926,
238
+ "smooth_aux_loss": -0.006348234789744461,
239
+ "total_training_time": 490.9107460975647
240
+ }
241
+ }