Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
68b1afd
·
verified ·
1 Parent(s): 6d7793e

Upload 2rva_2gpt_gs_8kcodes_i41yxf1l/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. 2rva_2gpt_gs_8kcodes_i41yxf1l/meta_002870.json +245 -0
2rva_2gpt_gs_8kcodes_i41yxf1l/meta_002870.json ADDED
@@ -0,0 +1,245 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": 4.200659059362183,
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": 6,
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
+ 3
22
+ ],
23
+ "mamba_blocks": [],
24
+ "gdn_blocks": [],
25
+ "gla_blocks": [],
26
+ "moe_blocks": [],
27
+ "kla_blocks": [],
28
+ "recurrent_vocab_sizes": [
29
+ 265,
30
+ 8192,
31
+ 8192,
32
+ 8192
33
+ ],
34
+ "rva_balance_loss_weight": 0.0,
35
+ "rva_balance_loss_temperature": 1.0,
36
+ "rva_entropy_loss_weight": 0.0,
37
+ "rva_use_ema": true,
38
+ "rva_ema_decay": 0.99,
39
+ "rva_ema_epsilon": 1e-05,
40
+ "rva_use_random_reset": false,
41
+ "rva_reset_threshold": 1.0,
42
+ "rva_reset_prob": 1.0,
43
+ "rva_use_l2_norm": false,
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
+ "rva_n_embd": 0,
58
+ "rva_no_mlp": false,
59
+ "rva_unified_embed": false,
60
+ "rva_embed_in_muon": true,
61
+ "rva_pq_groups": 0,
62
+ "rva_pq_k_sub": 0,
63
+ "rva_pq_balance_loss_weight": 0.0,
64
+ "rva_use_gumbel": true,
65
+ "rva_gumbel_temp": 2.0,
66
+ "rva_gumbel_temp_min": 0.1,
67
+ "rva_gumbel_anneal_steps": 0,
68
+ "rva_gumbel_diversity_weight": 0.001,
69
+ "rva_factored": false,
70
+ "rva_sub_vocab_size": 0,
71
+ "use_skip_connection": true,
72
+ "clustering_init_num_samples": 0,
73
+ "max_active_tokens_in_batch": -1,
74
+ "max_active_token_fraction": 0.0,
75
+ "d_state": 16,
76
+ "mamba_params": true,
77
+ "kla_kernel": true,
78
+ "mimo_rank": 1,
79
+ "skip_around_kla": true,
80
+ "decoder_mlp": false,
81
+ "use_reparametrisation_trick": true,
82
+ "num_experts": 4,
83
+ "moe_top_k": 2,
84
+ "moe_capacity_factor": 1.0,
85
+ "moe_aux_loss_weight": 0.01,
86
+ "moe_embedding_experts": false,
87
+ "num_iterations": 2870,
88
+ "target_flops": -1.0,
89
+ "target_param_data_ratio": 20,
90
+ "data_dir": "base_data",
91
+ "device_batch_size": 8,
92
+ "total_batch_size": 524288,
93
+ "use_muon": true,
94
+ "embedding_lr": 0.3,
95
+ "unembedding_lr": 0.004,
96
+ "grad_clip": 1.0,
97
+ "weight_decay": 0.0,
98
+ "matrix_lr": 0.02,
99
+ "adam_beta1": 0.8,
100
+ "adam_beta2": 0.95,
101
+ "warmup_ratio": 0.05,
102
+ "warmdown_ratio": 0.4,
103
+ "final_lr_frac": 0.0,
104
+ "resume_from_step": -1,
105
+ "eval_every": 250,
106
+ "eval_tokens": 10485760,
107
+ "core_metric_every": -1,
108
+ "core_metric_max_per_task": 500,
109
+ "sample_every": 250,
110
+ "save_every": 1000,
111
+ "push_checkpoints_to_hub": true,
112
+ "profile_rva_codebook_usage": true,
113
+ "model_tag": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
114
+ "stage": null,
115
+ "n_layer": 6,
116
+ "n_head": 4,
117
+ "n_kv_head": 4,
118
+ "n_embd": 256
119
+ },
120
+ "user_config": {
121
+ "run": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
122
+ "wandb_group": "gumbel_ablation",
123
+ "seed": 42,
124
+ "device_type": "cuda",
125
+ "config": null,
126
+ "depth": 6,
127
+ "aspect_ratio": 64,
128
+ "model_dim": 256,
129
+ "head_dim": 64,
130
+ "sequence_len": 1024,
131
+ "prefill_len": 0,
132
+ "vocab_size": 265,
133
+ "rva_blocks": [
134
+ 0,
135
+ 1,
136
+ 2,
137
+ 3
138
+ ],
139
+ "mamba_blocks": [],
140
+ "gdn_blocks": [],
141
+ "gla_blocks": [],
142
+ "moe_blocks": [],
143
+ "kla_blocks": [],
144
+ "recurrent_vocab_sizes": [
145
+ 265,
146
+ 8192,
147
+ 8192,
148
+ 8192
149
+ ],
150
+ "rva_balance_loss_weight": 0.0,
151
+ "rva_balance_loss_temperature": 1.0,
152
+ "rva_entropy_loss_weight": 0.0,
153
+ "rva_use_ema": true,
154
+ "rva_ema_decay": 0.99,
155
+ "rva_ema_epsilon": 1e-05,
156
+ "rva_use_random_reset": false,
157
+ "rva_reset_threshold": 1.0,
158
+ "rva_reset_prob": 1.0,
159
+ "rva_use_l2_norm": false,
160
+ "rva_use_fsq_ae": false,
161
+ "rva_fsq_levels": [
162
+ 4,
163
+ 4,
164
+ 4,
165
+ 4,
166
+ 4,
167
+ 4,
168
+ 4
169
+ ],
170
+ "rva_recon_loss_init": 1.0,
171
+ "rva_recon_loss_final": 0.01,
172
+ "rva_recon_anneal_fraction": 0.8,
173
+ "rva_n_embd": 0,
174
+ "rva_no_mlp": false,
175
+ "rva_unified_embed": false,
176
+ "rva_embed_in_muon": true,
177
+ "rva_pq_groups": 0,
178
+ "rva_pq_k_sub": 0,
179
+ "rva_pq_balance_loss_weight": 0.0,
180
+ "rva_use_gumbel": true,
181
+ "rva_gumbel_temp": 2.0,
182
+ "rva_gumbel_temp_min": 0.1,
183
+ "rva_gumbel_anneal_steps": 0,
184
+ "rva_gumbel_diversity_weight": 0.001,
185
+ "rva_factored": false,
186
+ "rva_sub_vocab_size": 0,
187
+ "use_skip_connection": true,
188
+ "clustering_init_num_samples": 0,
189
+ "max_active_tokens_in_batch": -1,
190
+ "max_active_token_fraction": 0.0,
191
+ "d_state": 16,
192
+ "mamba_params": true,
193
+ "kla_kernel": true,
194
+ "mimo_rank": 1,
195
+ "skip_around_kla": true,
196
+ "decoder_mlp": false,
197
+ "use_reparametrisation_trick": true,
198
+ "num_experts": 4,
199
+ "moe_top_k": 2,
200
+ "moe_capacity_factor": 1.0,
201
+ "moe_aux_loss_weight": 0.01,
202
+ "moe_embedding_experts": false,
203
+ "num_iterations": 2870,
204
+ "target_flops": -1.0,
205
+ "target_param_data_ratio": 20,
206
+ "data_dir": "base_data",
207
+ "device_batch_size": 8,
208
+ "total_batch_size": 524288,
209
+ "use_muon": true,
210
+ "embedding_lr": 0.3,
211
+ "unembedding_lr": 0.004,
212
+ "grad_clip": 1.0,
213
+ "weight_decay": 0.0,
214
+ "matrix_lr": 0.02,
215
+ "adam_beta1": 0.8,
216
+ "adam_beta2": 0.95,
217
+ "warmup_ratio": 0.05,
218
+ "warmdown_ratio": 0.4,
219
+ "final_lr_frac": 0.0,
220
+ "resume_from_step": -1,
221
+ "eval_every": 250,
222
+ "eval_tokens": 10485760,
223
+ "core_metric_every": -1,
224
+ "core_metric_max_per_task": 500,
225
+ "sample_every": 250,
226
+ "save_every": 1000,
227
+ "push_checkpoints_to_hub": true,
228
+ "profile_rva_codebook_usage": true,
229
+ "model_tag": "gumbel_ablation/2rva_2gpt_gs_8kcodes",
230
+ "stage": null
231
+ },
232
+ "stage": null,
233
+ "device_batch_size": 8,
234
+ "sequence_len": 1024,
235
+ "dataloader_state_dict": {
236
+ "pq_idx": 5,
237
+ "rg_idx": 32
238
+ },
239
+ "loop_state": {
240
+ "min_val_bpb": 4.200344690419538,
241
+ "smooth_train_loss": 2.8583947208189686,
242
+ "smooth_aux_loss": -0.023205555390662085,
243
+ "total_training_time": 2551.8576588630676
244
+ }
245
+ }