Hugging Face's logo

ASzecsenyi
/
VQLM

Model card Files
xet
 Community
ASzecsenyi commited on
Commit
d3258ff
·
verified ·
1 Parent(s): 849b4be

Upload gpt_2l_z4kucirm/meta_002870.json with huggingface_hub

Browse files
Files changed (1) hide show
  1. gpt_2l_z4kucirm/meta_002870.json +215 -0
gpt_2l_z4kucirm/meta_002870.json ADDED
@@ -0,0 +1,215 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "step": 2870,
3
+ "val_bpb": 1.4191577597040417,
4
+ "model_config": {
5
+ "run": "rva_ablation/gpt_2l",
6
+ "wandb_group": "rva_ablation",
7
+ "seed": 42,
8
+ "device_type": "cuda",
9
+ "config": null,
10
+ "depth": 2,
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
+ "mamba_blocks": [],
19
+ "gdn_blocks": [],
20
+ "gla_blocks": [],
21
+ "moe_blocks": [],
22
+ "kla_blocks": [],
23
+ "recurrent_vocab_sizes": [],
24
+ "rva_balance_loss_weight": 0.0,
25
+ "rva_balance_loss_temperature": 1.0,
26
+ "rva_entropy_loss_weight": 0.0,
27
+ "rva_use_ema": true,
28
+ "rva_ema_decay": 0.99,
29
+ "rva_ema_epsilon": 1e-05,
30
+ "rva_use_random_reset": false,
31
+ "rva_reset_threshold": 1.0,
32
+ "rva_reset_prob": 1.0,
33
+ "rva_use_l2_norm": false,
34
+ "rva_use_fsq_ae": false,
35
+ "rva_fsq_levels": [
36
+ 4,
37
+ 4,
38
+ 4,
39
+ 4,
40
+ 4,
41
+ 4,
42
+ 4
43
+ ],
44
+ "rva_recon_loss_init": 1.0,
45
+ "rva_recon_loss_final": 0.01,
46
+ "rva_recon_anneal_fraction": 0.8,
47
+ "rva_n_embd": 0,
48
+ "rva_no_mlp": false,
49
+ "rva_unified_embed": false,
50
+ "rva_embed_in_muon": true,
51
+ "rva_pq_groups": 0,
52
+ "rva_pq_k_sub": 0,
53
+ "rva_pq_balance_loss_weight": 0.0,
54
+ "rva_factored": false,
55
+ "rva_sub_vocab_size": 0,
56
+ "use_skip_connection": true,
57
+ "clustering_init_num_samples": 10000,
58
+ "max_active_tokens_in_batch": -1,
59
+ "max_active_token_fraction": 0.0,
60
+ "d_state": 16,
61
+ "mamba_params": true,
62
+ "kla_kernel": true,
63
+ "mimo_rank": 1,
64
+ "skip_around_kla": true,
65
+ "decoder_mlp": false,
66
+ "use_reparametrisation_trick": true,
67
+ "num_experts": 4,
68
+ "moe_top_k": 2,
69
+ "moe_capacity_factor": 1.0,
70
+ "moe_aux_loss_weight": 0.01,
71
+ "moe_embedding_experts": false,
72
+ "num_iterations": 2870,
73
+ "target_flops": -1.0,
74
+ "target_param_data_ratio": 20,
75
+ "data_dir": "base_data",
76
+ "device_batch_size": 64,
77
+ "total_batch_size": 524288,
78
+ "use_muon": true,
79
+ "embedding_lr": 0.3,
80
+ "unembedding_lr": 0.004,
81
+ "grad_clip": 1.0,
82
+ "weight_decay": 0.0,
83
+ "matrix_lr": 0.02,
84
+ "adam_beta1": 0.8,
85
+ "adam_beta2": 0.95,
86
+ "warmup_ratio": 0.05,
87
+ "warmdown_ratio": 0.4,
88
+ "final_lr_frac": 0.0,
89
+ "resume_from_step": -1,
90
+ "eval_every": 250,
91
+ "eval_tokens": 10485760,
92
+ "core_metric_every": -1,
93
+ "core_metric_max_per_task": 500,
94
+ "sample_every": 250,
95
+ "save_every": 1000,
96
+ "push_checkpoints_to_hub": true,
97
+ "profile_rva_codebook_usage": true,
98
+ "model_tag": "rva_ablation/gpt_2l",
99
+ "stage": null,
100
+ "n_layer": 2,
101
+ "n_head": 4,
102
+ "n_kv_head": 4,
103
+ "n_embd": 256
104
+ },
105
+ "user_config": {
106
+ "run": "rva_ablation/gpt_2l",
107
+ "wandb_group": "rva_ablation",
108
+ "seed": 42,
109
+ "device_type": "cuda",
110
+ "config": null,
111
+ "depth": 2,
112
+ "aspect_ratio": 64,
113
+ "model_dim": 256,
114
+ "head_dim": 64,
115
+ "sequence_len": 1024,
116
+ "prefill_len": 0,
117
+ "vocab_size": 1024,
118
+ "rva_blocks": [],
119
+ "mamba_blocks": [],
120
+ "gdn_blocks": [],
121
+ "gla_blocks": [],
122
+ "moe_blocks": [],
123
+ "kla_blocks": [],
124
+ "recurrent_vocab_sizes": [],
125
+ "rva_balance_loss_weight": 0.0,
126
+ "rva_balance_loss_temperature": 1.0,
127
+ "rva_entropy_loss_weight": 0.0,
128
+ "rva_use_ema": true,
129
+ "rva_ema_decay": 0.99,
130
+ "rva_ema_epsilon": 1e-05,
131
+ "rva_use_random_reset": false,
132
+ "rva_reset_threshold": 1.0,
133
+ "rva_reset_prob": 1.0,
134
+ "rva_use_l2_norm": false,
135
+ "rva_use_fsq_ae": false,
136
+ "rva_fsq_levels": [
137
+ 4,
138
+ 4,
139
+ 4,
140
+ 4,
141
+ 4,
142
+ 4,
143
+ 4
144
+ ],
145
+ "rva_recon_loss_init": 1.0,
146
+ "rva_recon_loss_final": 0.01,
147
+ "rva_recon_anneal_fraction": 0.8,
148
+ "rva_n_embd": 0,
149
+ "rva_no_mlp": false,
150
+ "rva_unified_embed": false,
151
+ "rva_embed_in_muon": true,
152
+ "rva_pq_groups": 0,
153
+ "rva_pq_k_sub": 0,
154
+ "rva_pq_balance_loss_weight": 0.0,
155
+ "rva_factored": false,
156
+ "rva_sub_vocab_size": 0,
157
+ "use_skip_connection": true,
158
+ "clustering_init_num_samples": 10000,
159
+ "max_active_tokens_in_batch": -1,
160
+ "max_active_token_fraction": 0.0,
161
+ "d_state": 16,
162
+ "mamba_params": true,
163
+ "kla_kernel": true,
164
+ "mimo_rank": 1,
165
+ "skip_around_kla": true,
166
+ "decoder_mlp": false,
167
+ "use_reparametrisation_trick": true,
168
+ "num_experts": 4,
169
+ "moe_top_k": 2,
170
+ "moe_capacity_factor": 1.0,
171
+ "moe_aux_loss_weight": 0.01,
172
+ "moe_embedding_experts": false,
173
+ "num_iterations": 2870,
174
+ "target_flops": -1.0,
175
+ "target_param_data_ratio": 20,
176
+ "data_dir": "base_data",
177
+ "device_batch_size": 64,
178
+ "total_batch_size": 524288,
179
+ "use_muon": true,
180
+ "embedding_lr": 0.3,
181
+ "unembedding_lr": 0.004,
182
+ "grad_clip": 1.0,
183
+ "weight_decay": 0.0,
184
+ "matrix_lr": 0.02,
185
+ "adam_beta1": 0.8,
186
+ "adam_beta2": 0.95,
187
+ "warmup_ratio": 0.05,
188
+ "warmdown_ratio": 0.4,
189
+ "final_lr_frac": 0.0,
190
+ "resume_from_step": -1,
191
+ "eval_every": 250,
192
+ "eval_tokens": 10485760,
193
+ "core_metric_every": -1,
194
+ "core_metric_max_per_task": 500,
195
+ "sample_every": 250,
196
+ "save_every": 1000,
197
+ "push_checkpoints_to_hub": true,
198
+ "profile_rva_codebook_usage": true,
199
+ "model_tag": "rva_ablation/gpt_2l",
200
+ "stage": null
201
+ },
202
+ "stage": null,
203
+ "device_batch_size": 64,
204
+ "sequence_len": 1024,
205
+ "dataloader_state_dict": {
206
+ "pq_idx": 13,
207
+ "rg_idx": 48
208
+ },
209
+ "loop_state": {
210
+ "min_val_bpb": 1.4191577597040417,
211
+ "smooth_train_loss": 2.504561127147281,
212
+ "smooth_aux_loss": 0.0,
213
+ "total_training_time": 278.20582270622253
214
+ }
215
+ }