Upload 2rva_2gpt_fsq_512_l3_r4cxvpxg/meta_001000.json with huggingface_hub
Browse files
2rva_2gpt_fsq_512_l3_r4cxvpxg/meta_001000.json
ADDED
|
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 1000,
|
| 3 |
+
"val_bpb": 2.0285634155579633,
|
| 4 |
+
"model_config": {
|
| 5 |
+
"run": "fsq_ablation/2rva_2gpt_fsq_512_l3",
|
| 6 |
+
"wandb_group": "fsq_ablation",
|
| 7 |
+
"seed": 42,
|
| 8 |
+
"device_type": "cuda",
|
| 9 |
+
"config": null,
|
| 10 |
+
"depth": 4,
|
| 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 |
+
],
|
| 21 |
+
"mamba_blocks": [],
|
| 22 |
+
"gdn_blocks": [],
|
| 23 |
+
"gla_blocks": [],
|
| 24 |
+
"moe_blocks": [],
|
| 25 |
+
"kla_blocks": [],
|
| 26 |
+
"recurrent_vocab_sizes": [
|
| 27 |
+
265,
|
| 28 |
+
512
|
| 29 |
+
],
|
| 30 |
+
"rva_balance_loss_weight": 0.0,
|
| 31 |
+
"rva_balance_loss_temperature": 1.0,
|
| 32 |
+
"rva_entropy_loss_weight": 0.0,
|
| 33 |
+
"rva_use_ema": true,
|
| 34 |
+
"rva_ema_decay": 0.99,
|
| 35 |
+
"rva_ema_epsilon": 1e-05,
|
| 36 |
+
"rva_use_random_reset": false,
|
| 37 |
+
"rva_reset_threshold": 1.0,
|
| 38 |
+
"rva_reset_prob": 1.0,
|
| 39 |
+
"rva_use_l2_norm": false,
|
| 40 |
+
"rva_use_fsq_ae": true,
|
| 41 |
+
"rva_fsq_levels": [
|
| 42 |
+
8,
|
| 43 |
+
8,
|
| 44 |
+
8
|
| 45 |
+
],
|
| 46 |
+
"rva_recon_loss_init": 1.0,
|
| 47 |
+
"rva_recon_loss_final": 0.01,
|
| 48 |
+
"rva_recon_anneal_fraction": 0.8,
|
| 49 |
+
"rva_n_embd": 0,
|
| 50 |
+
"rva_no_mlp": false,
|
| 51 |
+
"rva_unified_embed": false,
|
| 52 |
+
"rva_embed_in_muon": true,
|
| 53 |
+
"rva_fsq_adamw": false,
|
| 54 |
+
"rva_fsq_no_prenorm": false,
|
| 55 |
+
"rva_pq_groups": 0,
|
| 56 |
+
"rva_pq_k_sub": 0,
|
| 57 |
+
"rva_pq_balance_loss_weight": 0.0,
|
| 58 |
+
"rva_use_gumbel": false,
|
| 59 |
+
"rva_gumbel_temp": 2.0,
|
| 60 |
+
"rva_gumbel_temp_min": 0.5,
|
| 61 |
+
"rva_gumbel_anneal_steps": 0,
|
| 62 |
+
"rva_gumbel_diversity_weight": 0.1,
|
| 63 |
+
"rva_factored": false,
|
| 64 |
+
"rva_sub_vocab_size": 0,
|
| 65 |
+
"use_skip_connection": true,
|
| 66 |
+
"clustering_init_num_samples": 0,
|
| 67 |
+
"max_active_tokens_in_batch": -1,
|
| 68 |
+
"max_active_token_fraction": 0.0,
|
| 69 |
+
"d_state": 16,
|
| 70 |
+
"mamba_params": true,
|
| 71 |
+
"kla_kernel": true,
|
| 72 |
+
"mimo_rank": 1,
|
| 73 |
+
"skip_around_kla": true,
|
| 74 |
+
"decoder_mlp": false,
|
| 75 |
+
"use_reparametrisation_trick": true,
|
| 76 |
+
"num_experts": 4,
|
| 77 |
+
"moe_top_k": 2,
|
| 78 |
+
"moe_capacity_factor": 1.0,
|
| 79 |
+
"moe_aux_loss_weight": 0.01,
|
| 80 |
+
"moe_embedding_experts": false,
|
| 81 |
+
"num_iterations": 2870,
|
| 82 |
+
"target_flops": -1.0,
|
| 83 |
+
"target_param_data_ratio": 20,
|
| 84 |
+
"data_dir": "base_data",
|
| 85 |
+
"device_batch_size": 64,
|
| 86 |
+
"total_batch_size": 524288,
|
| 87 |
+
"use_muon": true,
|
| 88 |
+
"embedding_lr": 0.3,
|
| 89 |
+
"unembedding_lr": 0.004,
|
| 90 |
+
"grad_clip": 1.0,
|
| 91 |
+
"weight_decay": 0.0,
|
| 92 |
+
"matrix_lr": 0.02,
|
| 93 |
+
"adam_beta1": 0.8,
|
| 94 |
+
"adam_beta2": 0.95,
|
| 95 |
+
"warmup_ratio": 0.05,
|
| 96 |
+
"warmdown_ratio": 0.4,
|
| 97 |
+
"final_lr_frac": 0.0,
|
| 98 |
+
"resume_from_step": -1,
|
| 99 |
+
"eval_every": 250,
|
| 100 |
+
"eval_tokens": 10485760,
|
| 101 |
+
"core_metric_every": -1,
|
| 102 |
+
"core_metric_max_per_task": 500,
|
| 103 |
+
"sample_every": 250,
|
| 104 |
+
"save_every": 1000,
|
| 105 |
+
"push_checkpoints_to_hub": true,
|
| 106 |
+
"profile_rva_codebook_usage": true,
|
| 107 |
+
"model_tag": "fsq_ablation/2rva_2gpt_fsq_512_l3",
|
| 108 |
+
"stage": null,
|
| 109 |
+
"n_layer": 4,
|
| 110 |
+
"n_head": 4,
|
| 111 |
+
"n_kv_head": 4,
|
| 112 |
+
"n_embd": 256
|
| 113 |
+
},
|
| 114 |
+
"user_config": {
|
| 115 |
+
"run": "fsq_ablation/2rva_2gpt_fsq_512_l3",
|
| 116 |
+
"wandb_group": "fsq_ablation",
|
| 117 |
+
"seed": 42,
|
| 118 |
+
"device_type": "cuda",
|
| 119 |
+
"config": null,
|
| 120 |
+
"depth": 4,
|
| 121 |
+
"aspect_ratio": 64,
|
| 122 |
+
"model_dim": 256,
|
| 123 |
+
"head_dim": 64,
|
| 124 |
+
"sequence_len": 1024,
|
| 125 |
+
"prefill_len": 0,
|
| 126 |
+
"vocab_size": 265,
|
| 127 |
+
"rva_blocks": [
|
| 128 |
+
0,
|
| 129 |
+
1
|
| 130 |
+
],
|
| 131 |
+
"mamba_blocks": [],
|
| 132 |
+
"gdn_blocks": [],
|
| 133 |
+
"gla_blocks": [],
|
| 134 |
+
"moe_blocks": [],
|
| 135 |
+
"kla_blocks": [],
|
| 136 |
+
"recurrent_vocab_sizes": [
|
| 137 |
+
265,
|
| 138 |
+
512
|
| 139 |
+
],
|
| 140 |
+
"rva_balance_loss_weight": 0.0,
|
| 141 |
+
"rva_balance_loss_temperature": 1.0,
|
| 142 |
+
"rva_entropy_loss_weight": 0.0,
|
| 143 |
+
"rva_use_ema": true,
|
| 144 |
+
"rva_ema_decay": 0.99,
|
| 145 |
+
"rva_ema_epsilon": 1e-05,
|
| 146 |
+
"rva_use_random_reset": false,
|
| 147 |
+
"rva_reset_threshold": 1.0,
|
| 148 |
+
"rva_reset_prob": 1.0,
|
| 149 |
+
"rva_use_l2_norm": false,
|
| 150 |
+
"rva_use_fsq_ae": true,
|
| 151 |
+
"rva_fsq_levels": [
|
| 152 |
+
8,
|
| 153 |
+
8,
|
| 154 |
+
8
|
| 155 |
+
],
|
| 156 |
+
"rva_recon_loss_init": 1.0,
|
| 157 |
+
"rva_recon_loss_final": 0.01,
|
| 158 |
+
"rva_recon_anneal_fraction": 0.8,
|
| 159 |
+
"rva_n_embd": 0,
|
| 160 |
+
"rva_no_mlp": false,
|
| 161 |
+
"rva_unified_embed": false,
|
| 162 |
+
"rva_embed_in_muon": true,
|
| 163 |
+
"rva_fsq_adamw": false,
|
| 164 |
+
"rva_fsq_no_prenorm": false,
|
| 165 |
+
"rva_pq_groups": 0,
|
| 166 |
+
"rva_pq_k_sub": 0,
|
| 167 |
+
"rva_pq_balance_loss_weight": 0.0,
|
| 168 |
+
"rva_use_gumbel": false,
|
| 169 |
+
"rva_gumbel_temp": 2.0,
|
| 170 |
+
"rva_gumbel_temp_min": 0.5,
|
| 171 |
+
"rva_gumbel_anneal_steps": 0,
|
| 172 |
+
"rva_gumbel_diversity_weight": 0.1,
|
| 173 |
+
"rva_factored": false,
|
| 174 |
+
"rva_sub_vocab_size": 0,
|
| 175 |
+
"use_skip_connection": true,
|
| 176 |
+
"clustering_init_num_samples": 0,
|
| 177 |
+
"max_active_tokens_in_batch": -1,
|
| 178 |
+
"max_active_token_fraction": 0.0,
|
| 179 |
+
"d_state": 16,
|
| 180 |
+
"mamba_params": true,
|
| 181 |
+
"kla_kernel": true,
|
| 182 |
+
"mimo_rank": 1,
|
| 183 |
+
"skip_around_kla": true,
|
| 184 |
+
"decoder_mlp": false,
|
| 185 |
+
"use_reparametrisation_trick": true,
|
| 186 |
+
"num_experts": 4,
|
| 187 |
+
"moe_top_k": 2,
|
| 188 |
+
"moe_capacity_factor": 1.0,
|
| 189 |
+
"moe_aux_loss_weight": 0.01,
|
| 190 |
+
"moe_embedding_experts": false,
|
| 191 |
+
"num_iterations": 2870,
|
| 192 |
+
"target_flops": -1.0,
|
| 193 |
+
"target_param_data_ratio": 20,
|
| 194 |
+
"data_dir": "base_data",
|
| 195 |
+
"device_batch_size": 64,
|
| 196 |
+
"total_batch_size": 524288,
|
| 197 |
+
"use_muon": true,
|
| 198 |
+
"embedding_lr": 0.3,
|
| 199 |
+
"unembedding_lr": 0.004,
|
| 200 |
+
"grad_clip": 1.0,
|
| 201 |
+
"weight_decay": 0.0,
|
| 202 |
+
"matrix_lr": 0.02,
|
| 203 |
+
"adam_beta1": 0.8,
|
| 204 |
+
"adam_beta2": 0.95,
|
| 205 |
+
"warmup_ratio": 0.05,
|
| 206 |
+
"warmdown_ratio": 0.4,
|
| 207 |
+
"final_lr_frac": 0.0,
|
| 208 |
+
"resume_from_step": -1,
|
| 209 |
+
"eval_every": 250,
|
| 210 |
+
"eval_tokens": 10485760,
|
| 211 |
+
"core_metric_every": -1,
|
| 212 |
+
"core_metric_max_per_task": 500,
|
| 213 |
+
"sample_every": 250,
|
| 214 |
+
"save_every": 1000,
|
| 215 |
+
"push_checkpoints_to_hub": true,
|
| 216 |
+
"profile_rva_codebook_usage": true,
|
| 217 |
+
"model_tag": "fsq_ablation/2rva_2gpt_fsq_512_l3",
|
| 218 |
+
"stage": null
|
| 219 |
+
},
|
| 220 |
+
"stage": null,
|
| 221 |
+
"device_batch_size": 64,
|
| 222 |
+
"sequence_len": 1024,
|
| 223 |
+
"dataloader_state_dict": {
|
| 224 |
+
"pq_idx": 1,
|
| 225 |
+
"rg_idx": 48
|
| 226 |
+
},
|
| 227 |
+
"loop_state": {
|
| 228 |
+
"min_val_bpb": 2.0285634155579633,
|
| 229 |
+
"smooth_train_loss": 1.430909804675896,
|
| 230 |
+
"smooth_aux_loss": 0.0,
|
| 231 |
+
"total_training_time": 132.36980772018433
|
| 232 |
+
}
|
| 233 |
+
}
|