Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
5af103e
·
verified ·
1 Parent(s): fb608c5

Upload rva_3x256_5bu1f8yt/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_3x256_5bu1f8yt/meta_002870.json +211 -0
rva_3x256_5bu1f8yt/meta_002870.json ADDED
@@ -0,0 +1,211 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": 1.5884641897160392,
4
+ "model_config": {
5
+ "run": "rva_3x256",
6
+ "wandb_group": null,
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": "rvamix_1l",
10
+ "depth": 3,
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
+ ],
20
+ "mamba_blocks": [],
21
+ "gdn_blocks": [],
22
+ "gla_blocks": [],
23
+ "moe_blocks": [],
24
+ "kla_blocks": [],
25
+ "recurrent_vocab_sizes": [
26
+ 265
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_unified_embed": false,
52
+ "rva_factored": true,
53
+ "rva_sub_vocab_size": 0,
54
+ "use_skip_connection": true,
55
+ "clustering_init_num_samples": 0,
56
+ "max_active_tokens_in_batch": -1,
57
+ "max_active_token_fraction": 0.0,
58
+ "d_state": 16,
59
+ "mamba_params": true,
60
+ "kla_kernel": true,
61
+ "mimo_rank": 1,
62
+ "skip_around_kla": true,
63
+ "decoder_mlp": false,
64
+ "use_reparametrisation_trick": true,
65
+ "num_experts": 4,
66
+ "moe_top_k": 2,
67
+ "moe_capacity_factor": 1.0,
68
+ "moe_aux_loss_weight": 0.01,
69
+ "moe_embedding_experts": false,
70
+ "num_iterations": 2870,
71
+ "target_flops": -1.0,
72
+ "target_param_data_ratio": 20,
73
+ "data_dir": "base_data",
74
+ "device_batch_size": 64,
75
+ "total_batch_size": 524288,
76
+ "use_muon": true,
77
+ "embedding_lr": 0.3,
78
+ "unembedding_lr": 0.004,
79
+ "grad_clip": 1.0,
80
+ "weight_decay": 0.0,
81
+ "matrix_lr": 0.02,
82
+ "adam_beta1": 0.8,
83
+ "adam_beta2": 0.95,
84
+ "warmup_ratio": 0.05,
85
+ "warmdown_ratio": 0.4,
86
+ "final_lr_frac": 0.0,
87
+ "resume_from_step": -1,
88
+ "eval_every": 250,
89
+ "eval_tokens": 10485760,
90
+ "core_metric_every": -1,
91
+ "core_metric_max_per_task": 500,
92
+ "sample_every": 250,
93
+ "save_every": 1000,
94
+ "push_checkpoints_to_hub": true,
95
+ "profile_rva_codebook_usage": true,
96
+ "model_tag": "rva_3x256",
97
+ "stage": null,
98
+ "n_layer": 3,
99
+ "n_head": 16,
100
+ "n_kv_head": 16,
101
+ "n_embd": 256
102
+ },
103
+ "user_config": {
104
+ "run": "rva_3x256",
105
+ "wandb_group": null,
106
+ "seed": 42,
107
+ "device_type": "cuda",
108
+ "config": "rvamix_1l",
109
+ "depth": 3,
110
+ "aspect_ratio": 64,
111
+ "model_dim": 256,
112
+ "head_dim": 16,
113
+ "sequence_len": 1024,
114
+ "prefill_len": 0,
115
+ "vocab_size": 265,
116
+ "rva_blocks": [
117
+ 0
118
+ ],
119
+ "mamba_blocks": [],
120
+ "gdn_blocks": [],
121
+ "gla_blocks": [],
122
+ "moe_blocks": [],
123
+ "kla_blocks": [],
124
+ "recurrent_vocab_sizes": [
125
+ 265
126
+ ],
127
+ "rva_balance_loss_weight": 0.0,
128
+ "rva_balance_loss_temperature": 1.0,
129
+ "rva_entropy_loss_weight": 0.0,
130
+ "rva_use_ema": true,
131
+ "rva_ema_decay": 0.99,
132
+ "rva_ema_epsilon": 1e-05,
133
+ "rva_use_random_reset": false,
134
+ "rva_reset_threshold": 1.0,
135
+ "rva_reset_prob": 1.0,
136
+ "rva_use_l2_norm": false,
137
+ "rva_use_fsq_ae": false,
138
+ "rva_fsq_levels": [
139
+ 4,
140
+ 4,
141
+ 4,
142
+ 4,
143
+ 4,
144
+ 4,
145
+ 4
146
+ ],
147
+ "rva_recon_loss_init": 1.0,
148
+ "rva_recon_loss_final": 0.01,
149
+ "rva_recon_anneal_fraction": 0.8,
150
+ "rva_unified_embed": false,
151
+ "rva_factored": true,
152
+ "rva_sub_vocab_size": 0,
153
+ "use_skip_connection": true,
154
+ "clustering_init_num_samples": 0,
155
+ "max_active_tokens_in_batch": -1,
156
+ "max_active_token_fraction": 0.0,
157
+ "d_state": 16,
158
+ "mamba_params": true,
159
+ "kla_kernel": true,
160
+ "mimo_rank": 1,
161
+ "skip_around_kla": true,
162
+ "decoder_mlp": false,
163
+ "use_reparametrisation_trick": true,
164
+ "num_experts": 4,
165
+ "moe_top_k": 2,
166
+ "moe_capacity_factor": 1.0,
167
+ "moe_aux_loss_weight": 0.01,
168
+ "moe_embedding_experts": false,
169
+ "num_iterations": 2870,
170
+ "target_flops": -1.0,
171
+ "target_param_data_ratio": 20,
172
+ "data_dir": "base_data",
173
+ "device_batch_size": 64,
174
+ "total_batch_size": 524288,
175
+ "use_muon": true,
176
+ "embedding_lr": 0.3,
177
+ "unembedding_lr": 0.004,
178
+ "grad_clip": 1.0,
179
+ "weight_decay": 0.0,
180
+ "matrix_lr": 0.02,
181
+ "adam_beta1": 0.8,
182
+ "adam_beta2": 0.95,
183
+ "warmup_ratio": 0.05,
184
+ "warmdown_ratio": 0.4,
185
+ "final_lr_frac": 0.0,
186
+ "resume_from_step": -1,
187
+ "eval_every": 250,
188
+ "eval_tokens": 10485760,
189
+ "core_metric_every": -1,
190
+ "core_metric_max_per_task": 500,
191
+ "sample_every": 250,
192
+ "save_every": 1000,
193
+ "push_checkpoints_to_hub": true,
194
+ "profile_rva_codebook_usage": true,
195
+ "model_tag": "rva_3x256",
196
+ "stage": null
197
+ },
198
+ "stage": null,
199
+ "device_batch_size": 64,
200
+ "sequence_len": 1024,
201
+ "dataloader_state_dict": {
202
+ "pq_idx": 5,
203
+ "rg_idx": 32
204
+ },
205
+ "loop_state": {
206
+ "min_val_bpb": 1.5884641897160392,
207
+ "smooth_train_loss": 1.0977336955298314,
208
+ "smooth_aux_loss": 0.0,
209
+ "total_training_time": 359.23615026474
210
+ }
211
+ }