Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
2f64d02
·
verified ·
1 Parent(s): 0072a09

Upload rva_2gpt_diag_1ajhrcun/meta_002000.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_2gpt_diag_1ajhrcun/meta_002000.json +223 -0
rva_2gpt_diag_1ajhrcun/meta_002000.json ADDED
@@ -0,0 +1,223 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2000,
3
+ "val_bpb": 1.4354157775496745,
4
+ "model_config": {
5
+ "run": "rva_ablation/rva_2gpt_diag",
6
+ "wandb_group": "rva_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_pq_groups": 0,
56
+ "rva_pq_k_sub": 0,
57
+ "rva_pq_balance_loss_weight": 0.0,
58
+ "rva_factored": true,
59
+ "rva_sub_vocab_size": 0,
60
+ "use_skip_connection": true,
61
+ "clustering_init_num_samples": 10000,
62
+ "max_active_tokens_in_batch": -1,
63
+ "max_active_token_fraction": 0.0,
64
+ "d_state": 16,
65
+ "mamba_params": true,
66
+ "kla_kernel": true,
67
+ "mimo_rank": 1,
68
+ "skip_around_kla": true,
69
+ "decoder_mlp": false,
70
+ "use_reparametrisation_trick": true,
71
+ "num_experts": 4,
72
+ "moe_top_k": 2,
73
+ "moe_capacity_factor": 1.0,
74
+ "moe_aux_loss_weight": 0.01,
75
+ "moe_embedding_experts": false,
76
+ "num_iterations": 2870,
77
+ "target_flops": -1.0,
78
+ "target_param_data_ratio": 20,
79
+ "data_dir": "base_data",
80
+ "device_batch_size": 64,
81
+ "total_batch_size": 524288,
82
+ "use_muon": true,
83
+ "embedding_lr": 0.3,
84
+ "unembedding_lr": 0.004,
85
+ "grad_clip": 1.0,
86
+ "weight_decay": 0.0,
87
+ "matrix_lr": 0.02,
88
+ "adam_beta1": 0.8,
89
+ "adam_beta2": 0.95,
90
+ "warmup_ratio": 0.05,
91
+ "warmdown_ratio": 0.4,
92
+ "final_lr_frac": 0.0,
93
+ "resume_from_step": -1,
94
+ "eval_every": 250,
95
+ "eval_tokens": 10485760,
96
+ "core_metric_every": -1,
97
+ "core_metric_max_per_task": 500,
98
+ "sample_every": 250,
99
+ "save_every": 1000,
100
+ "push_checkpoints_to_hub": true,
101
+ "profile_rva_codebook_usage": true,
102
+ "model_tag": "rva_ablation/rva_2gpt_diag",
103
+ "stage": null,
104
+ "n_layer": 3,
105
+ "n_head": 4,
106
+ "n_kv_head": 4,
107
+ "n_embd": 256
108
+ },
109
+ "user_config": {
110
+ "run": "rva_ablation/rva_2gpt_diag",
111
+ "wandb_group": "rva_ablation",
112
+ "seed": 42,
113
+ "device_type": "cuda",
114
+ "config": null,
115
+ "depth": 3,
116
+ "aspect_ratio": 64,
117
+ "model_dim": 256,
118
+ "head_dim": 64,
119
+ "sequence_len": 1024,
120
+ "prefill_len": 0,
121
+ "vocab_size": 1024,
122
+ "rva_blocks": [
123
+ 0
124
+ ],
125
+ "mamba_blocks": [],
126
+ "gdn_blocks": [],
127
+ "gla_blocks": [],
128
+ "moe_blocks": [],
129
+ "kla_blocks": [],
130
+ "recurrent_vocab_sizes": [
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_embed_in_muon": true,
160
+ "rva_pq_groups": 0,
161
+ "rva_pq_k_sub": 0,
162
+ "rva_pq_balance_loss_weight": 0.0,
163
+ "rva_factored": true,
164
+ "rva_sub_vocab_size": 0,
165
+ "use_skip_connection": true,
166
+ "clustering_init_num_samples": 10000,
167
+ "max_active_tokens_in_batch": -1,
168
+ "max_active_token_fraction": 0.0,
169
+ "d_state": 16,
170
+ "mamba_params": true,
171
+ "kla_kernel": true,
172
+ "mimo_rank": 1,
173
+ "skip_around_kla": true,
174
+ "decoder_mlp": false,
175
+ "use_reparametrisation_trick": true,
176
+ "num_experts": 4,
177
+ "moe_top_k": 2,
178
+ "moe_capacity_factor": 1.0,
179
+ "moe_aux_loss_weight": 0.01,
180
+ "moe_embedding_experts": false,
181
+ "num_iterations": 2870,
182
+ "target_flops": -1.0,
183
+ "target_param_data_ratio": 20,
184
+ "data_dir": "base_data",
185
+ "device_batch_size": 64,
186
+ "total_batch_size": 524288,
187
+ "use_muon": true,
188
+ "embedding_lr": 0.3,
189
+ "unembedding_lr": 0.004,
190
+ "grad_clip": 1.0,
191
+ "weight_decay": 0.0,
192
+ "matrix_lr": 0.02,
193
+ "adam_beta1": 0.8,
194
+ "adam_beta2": 0.95,
195
+ "warmup_ratio": 0.05,
196
+ "warmdown_ratio": 0.4,
197
+ "final_lr_frac": 0.0,
198
+ "resume_from_step": -1,
199
+ "eval_every": 250,
200
+ "eval_tokens": 10485760,
201
+ "core_metric_every": -1,
202
+ "core_metric_max_per_task": 500,
203
+ "sample_every": 250,
204
+ "save_every": 1000,
205
+ "push_checkpoints_to_hub": true,
206
+ "profile_rva_codebook_usage": true,
207
+ "model_tag": "rva_ablation/rva_2gpt_diag",
208
+ "stage": null
209
+ },
210
+ "stage": null,
211
+ "device_batch_size": 64,
212
+ "sequence_len": 1024,
213
+ "dataloader_state_dict": {
214
+ "pq_idx": 9,
215
+ "rg_idx": 40
216
+ },
217
+ "loop_state": {
218
+ "min_val_bpb": 1.4354157775496745,
219
+ "smooth_train_loss": 2.5347942990269487,
220
+ "smooth_aux_loss": 0.0,
221
+ "total_training_time": 220.3431875705719
222
+ }
223
+ }