Upload moe_2x256_hx8z5pqf/meta_000186.json with huggingface_hub
Browse files
moe_2x256_hx8z5pqf/meta_000186.json
ADDED
|
@@ -0,0 +1,144 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"step": 186,
|
| 3 |
+
"val_bpb": 2.006805166997607,
|
| 4 |
+
"model_config": {
|
| 5 |
+
"run": "moe_2x256",
|
| 6 |
+
"wandb_group": null,
|
| 7 |
+
"device_type": "cuda",
|
| 8 |
+
"config": "moe20m",
|
| 9 |
+
"depth": 2,
|
| 10 |
+
"aspect_ratio": 64,
|
| 11 |
+
"model_dim": 256,
|
| 12 |
+
"head_dim": 128,
|
| 13 |
+
"sequence_len": 1024,
|
| 14 |
+
"vocab_size": 265,
|
| 15 |
+
"rva_blocks": [],
|
| 16 |
+
"recurrent_vocab_sizes": [],
|
| 17 |
+
"kla_blocks": [],
|
| 18 |
+
"mamba_blocks": [],
|
| 19 |
+
"gdn_blocks": [],
|
| 20 |
+
"gla_blocks": [],
|
| 21 |
+
"moe_blocks": [
|
| 22 |
+
-1
|
| 23 |
+
],
|
| 24 |
+
"d_state": 16,
|
| 25 |
+
"mamba_params": true,
|
| 26 |
+
"kla_kernel": true,
|
| 27 |
+
"mimo_rank": 1,
|
| 28 |
+
"skip_around_kla": true,
|
| 29 |
+
"decoder_mlp": false,
|
| 30 |
+
"use_reparametrisation_trick": true,
|
| 31 |
+
"num_experts": 4,
|
| 32 |
+
"moe_top_k": 2,
|
| 33 |
+
"moe_capacity_factor": 1.0,
|
| 34 |
+
"moe_aux_loss_weight": 0.01,
|
| 35 |
+
"num_iterations": -1,
|
| 36 |
+
"target_flops": -1.0,
|
| 37 |
+
"target_param_data_ratio": 20,
|
| 38 |
+
"data_dir": "base_data",
|
| 39 |
+
"device_batch_size": 64,
|
| 40 |
+
"total_batch_size": 524288,
|
| 41 |
+
"use_muon": true,
|
| 42 |
+
"embedding_lr": 0.3,
|
| 43 |
+
"unembedding_lr": 0.004,
|
| 44 |
+
"grad_clip": 1.0,
|
| 45 |
+
"weight_decay": 0.0,
|
| 46 |
+
"matrix_lr": 0.02,
|
| 47 |
+
"adam_beta1": 0.8,
|
| 48 |
+
"adam_beta2": 0.95,
|
| 49 |
+
"warmup_ratio": 0.05,
|
| 50 |
+
"warmdown_ratio": 0.4,
|
| 51 |
+
"final_lr_frac": 0.0,
|
| 52 |
+
"resume_from_step": -1,
|
| 53 |
+
"eval_every": 250,
|
| 54 |
+
"eval_tokens": 10485760,
|
| 55 |
+
"core_metric_every": -1,
|
| 56 |
+
"core_metric_max_per_task": 500,
|
| 57 |
+
"sample_every": 250,
|
| 58 |
+
"save_every": 1000,
|
| 59 |
+
"push_checkpoints_to_hub": true,
|
| 60 |
+
"use_profiler": false,
|
| 61 |
+
"profile_step": 2,
|
| 62 |
+
"profile_micro_step": 0,
|
| 63 |
+
"memory_history_max_entries": 10000,
|
| 64 |
+
"model_tag": "moe_2x256",
|
| 65 |
+
"n_layer": 2,
|
| 66 |
+
"n_head": 2,
|
| 67 |
+
"n_kv_head": 2,
|
| 68 |
+
"n_embd": 256
|
| 69 |
+
},
|
| 70 |
+
"user_config": {
|
| 71 |
+
"run": "moe_2x256",
|
| 72 |
+
"wandb_group": null,
|
| 73 |
+
"device_type": "cuda",
|
| 74 |
+
"config": "moe20m",
|
| 75 |
+
"depth": 2,
|
| 76 |
+
"aspect_ratio": 64,
|
| 77 |
+
"model_dim": 256,
|
| 78 |
+
"head_dim": 128,
|
| 79 |
+
"sequence_len": 1024,
|
| 80 |
+
"vocab_size": 265,
|
| 81 |
+
"rva_blocks": [],
|
| 82 |
+
"recurrent_vocab_sizes": [],
|
| 83 |
+
"kla_blocks": [],
|
| 84 |
+
"mamba_blocks": [],
|
| 85 |
+
"gdn_blocks": [],
|
| 86 |
+
"gla_blocks": [],
|
| 87 |
+
"moe_blocks": [
|
| 88 |
+
-1
|
| 89 |
+
],
|
| 90 |
+
"d_state": 16,
|
| 91 |
+
"mamba_params": true,
|
| 92 |
+
"kla_kernel": true,
|
| 93 |
+
"mimo_rank": 1,
|
| 94 |
+
"skip_around_kla": true,
|
| 95 |
+
"decoder_mlp": false,
|
| 96 |
+
"use_reparametrisation_trick": true,
|
| 97 |
+
"num_experts": 4,
|
| 98 |
+
"moe_top_k": 2,
|
| 99 |
+
"moe_capacity_factor": 1.0,
|
| 100 |
+
"moe_aux_loss_weight": 0.01,
|
| 101 |
+
"num_iterations": -1,
|
| 102 |
+
"target_flops": -1.0,
|
| 103 |
+
"target_param_data_ratio": 20,
|
| 104 |
+
"data_dir": "base_data",
|
| 105 |
+
"device_batch_size": 64,
|
| 106 |
+
"total_batch_size": 524288,
|
| 107 |
+
"use_muon": true,
|
| 108 |
+
"embedding_lr": 0.3,
|
| 109 |
+
"unembedding_lr": 0.004,
|
| 110 |
+
"grad_clip": 1.0,
|
| 111 |
+
"weight_decay": 0.0,
|
| 112 |
+
"matrix_lr": 0.02,
|
| 113 |
+
"adam_beta1": 0.8,
|
| 114 |
+
"adam_beta2": 0.95,
|
| 115 |
+
"warmup_ratio": 0.05,
|
| 116 |
+
"warmdown_ratio": 0.4,
|
| 117 |
+
"final_lr_frac": 0.0,
|
| 118 |
+
"resume_from_step": -1,
|
| 119 |
+
"eval_every": 250,
|
| 120 |
+
"eval_tokens": 10485760,
|
| 121 |
+
"core_metric_every": -1,
|
| 122 |
+
"core_metric_max_per_task": 500,
|
| 123 |
+
"sample_every": 250,
|
| 124 |
+
"save_every": 1000,
|
| 125 |
+
"push_checkpoints_to_hub": true,
|
| 126 |
+
"use_profiler": false,
|
| 127 |
+
"profile_step": 2,
|
| 128 |
+
"profile_micro_step": 0,
|
| 129 |
+
"memory_history_max_entries": 10000,
|
| 130 |
+
"model_tag": "moe_2x256"
|
| 131 |
+
},
|
| 132 |
+
"device_batch_size": 64,
|
| 133 |
+
"sequence_len": 1024,
|
| 134 |
+
"dataloader_state_dict": {
|
| 135 |
+
"pq_idx": 0,
|
| 136 |
+
"rg_idx": 20
|
| 137 |
+
},
|
| 138 |
+
"loop_state": {
|
| 139 |
+
"min_val_bpb": 2.006805166997607,
|
| 140 |
+
"smooth_train_loss": 1.4490783121314121,
|
| 141 |
+
"smooth_aux_loss": 0.020028265203088005,
|
| 142 |
+
"total_training_time": 17.731884717941284
|
| 143 |
+
}
|
| 144 |
+
}
|