Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
c6f2e9c
·
verified ·
1 Parent(s): 0ebe2eb

Upload rva_3x256_ydpzx20f/meta_001000.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. rva_3x256_ydpzx20f/meta_001000.json +202 -0
rva_3x256_ydpzx20f/meta_001000.json ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 1000,
3
+ "val_bpb": 1.7224356086046526,
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
+ "vocab_size": 265,
16
+ "rva_blocks": [
17
+ 0
18
+ ],
19
+ "mamba_blocks": [],
20
+ "gdn_blocks": [],
21
+ "gla_blocks": [],
22
+ "moe_blocks": [],
23
+ "kla_blocks": [],
24
+ "recurrent_vocab_sizes": [],
25
+ "rva_balance_loss_weight": 0.0,
26
+ "rva_balance_loss_temperature": 1.0,
27
+ "rva_entropy_loss_weight": 0.0,
28
+ "rva_use_ema": true,
29
+ "rva_ema_decay": 0.99,
30
+ "rva_ema_epsilon": 1e-05,
31
+ "rva_use_random_reset": false,
32
+ "rva_reset_threshold": 1.0,
33
+ "rva_reset_prob": 1.0,
34
+ "rva_use_l2_norm": false,
35
+ "rva_use_fsq_ae": false,
36
+ "rva_fsq_levels": [
37
+ 4,
38
+ 4,
39
+ 4,
40
+ 4,
41
+ 4,
42
+ 4,
43
+ 4
44
+ ],
45
+ "rva_recon_loss_init": 1.0,
46
+ "rva_recon_loss_final": 0.01,
47
+ "rva_recon_anneal_fraction": 0.8,
48
+ "rva_unified_embed": false,
49
+ "use_skip_connection": true,
50
+ "clustering_init_num_samples": 0,
51
+ "max_active_tokens_in_batch": -1,
52
+ "max_active_token_fraction": 0.0,
53
+ "d_state": 16,
54
+ "mamba_params": true,
55
+ "kla_kernel": true,
56
+ "mimo_rank": 1,
57
+ "skip_around_kla": true,
58
+ "decoder_mlp": false,
59
+ "use_reparametrisation_trick": true,
60
+ "num_experts": 4,
61
+ "moe_top_k": 2,
62
+ "moe_capacity_factor": 1.0,
63
+ "moe_aux_loss_weight": 0.01,
64
+ "moe_embedding_experts": false,
65
+ "num_iterations": 2870,
66
+ "target_flops": -1.0,
67
+ "target_param_data_ratio": 20,
68
+ "data_dir": "base_data",
69
+ "device_batch_size": 8,
70
+ "total_batch_size": 524288,
71
+ "use_muon": true,
72
+ "embedding_lr": 0.3,
73
+ "unembedding_lr": 0.004,
74
+ "grad_clip": 1.0,
75
+ "weight_decay": 0.0,
76
+ "matrix_lr": 0.02,
77
+ "adam_beta1": 0.8,
78
+ "adam_beta2": 0.95,
79
+ "warmup_ratio": 0.05,
80
+ "warmdown_ratio": 0.4,
81
+ "final_lr_frac": 0.0,
82
+ "resume_from_step": -1,
83
+ "eval_every": 250,
84
+ "eval_tokens": 10485760,
85
+ "core_metric_every": -1,
86
+ "core_metric_max_per_task": 500,
87
+ "sample_every": 250,
88
+ "save_every": 1000,
89
+ "push_checkpoints_to_hub": true,
90
+ "profile_rva_codebook_usage": true,
91
+ "model_tag": "rva_3x256",
92
+ "stage": null,
93
+ "n_layer": 3,
94
+ "n_head": 16,
95
+ "n_kv_head": 16,
96
+ "n_embd": 256
97
+ },
98
+ "user_config": {
99
+ "run": "rva_3x256",
100
+ "wandb_group": null,
101
+ "seed": 42,
102
+ "device_type": "cuda",
103
+ "config": "rvamix_1l",
104
+ "depth": 3,
105
+ "aspect_ratio": 64,
106
+ "model_dim": 256,
107
+ "head_dim": 16,
108
+ "sequence_len": 1024,
109
+ "vocab_size": 265,
110
+ "rva_blocks": [
111
+ 0
112
+ ],
113
+ "mamba_blocks": [],
114
+ "gdn_blocks": [],
115
+ "gla_blocks": [],
116
+ "moe_blocks": [],
117
+ "kla_blocks": [],
118
+ "recurrent_vocab_sizes": [],
119
+ "rva_balance_loss_weight": 0.0,
120
+ "rva_balance_loss_temperature": 1.0,
121
+ "rva_entropy_loss_weight": 0.0,
122
+ "rva_use_ema": true,
123
+ "rva_ema_decay": 0.99,
124
+ "rva_ema_epsilon": 1e-05,
125
+ "rva_use_random_reset": false,
126
+ "rva_reset_threshold": 1.0,
127
+ "rva_reset_prob": 1.0,
128
+ "rva_use_l2_norm": false,
129
+ "rva_use_fsq_ae": false,
130
+ "rva_fsq_levels": [
131
+ 4,
132
+ 4,
133
+ 4,
134
+ 4,
135
+ 4,
136
+ 4,
137
+ 4
138
+ ],
139
+ "rva_recon_loss_init": 1.0,
140
+ "rva_recon_loss_final": 0.01,
141
+ "rva_recon_anneal_fraction": 0.8,
142
+ "rva_unified_embed": false,
143
+ "use_skip_connection": true,
144
+ "clustering_init_num_samples": 0,
145
+ "max_active_tokens_in_batch": -1,
146
+ "max_active_token_fraction": 0.0,
147
+ "d_state": 16,
148
+ "mamba_params": true,
149
+ "kla_kernel": true,
150
+ "mimo_rank": 1,
151
+ "skip_around_kla": true,
152
+ "decoder_mlp": false,
153
+ "use_reparametrisation_trick": true,
154
+ "num_experts": 4,
155
+ "moe_top_k": 2,
156
+ "moe_capacity_factor": 1.0,
157
+ "moe_aux_loss_weight": 0.01,
158
+ "moe_embedding_experts": false,
159
+ "num_iterations": 2870,
160
+ "target_flops": -1.0,
161
+ "target_param_data_ratio": 20,
162
+ "data_dir": "base_data",
163
+ "device_batch_size": 8,
164
+ "total_batch_size": 524288,
165
+ "use_muon": true,
166
+ "embedding_lr": 0.3,
167
+ "unembedding_lr": 0.004,
168
+ "grad_clip": 1.0,
169
+ "weight_decay": 0.0,
170
+ "matrix_lr": 0.02,
171
+ "adam_beta1": 0.8,
172
+ "adam_beta2": 0.95,
173
+ "warmup_ratio": 0.05,
174
+ "warmdown_ratio": 0.4,
175
+ "final_lr_frac": 0.0,
176
+ "resume_from_step": -1,
177
+ "eval_every": 250,
178
+ "eval_tokens": 10485760,
179
+ "core_metric_every": -1,
180
+ "core_metric_max_per_task": 500,
181
+ "sample_every": 250,
182
+ "save_every": 1000,
183
+ "push_checkpoints_to_hub": true,
184
+ "profile_rva_codebook_usage": true,
185
+ "model_tag": "rva_3x256",
186
+ "stage": null,
187
+ "pod_name": "s2027538-infk8s-job-4bd5x-2bg7d"
188
+ },
189
+ "stage": null,
190
+ "device_batch_size": 8,
191
+ "sequence_len": 1024,
192
+ "dataloader_state_dict": {
193
+ "pq_idx": 2,
194
+ "rg_idx": 4
195
+ },
196
+ "loop_state": {
197
+ "min_val_bpb": 1.7224356086046526,
198
+ "smooth_train_loss": 1.215767225295908,
199
+ "smooth_aux_loss": 0.0,
200
+ "total_training_time": 559.3583030700684
201
+ }
202
+ }