Upload full_config.json with huggingface_hub
Browse files- full_config.json +73 -0
full_config.json
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"shorthand": "model.768.lyr.12 - seqlen.1024 - mla.0.128.0 - ah12.64 - rd32.32",
|
| 3 |
+
"notes": "GPT-2 run with only a kv space, increased to size 128.",
|
| 4 |
+
"model": {
|
| 5 |
+
"hidden_size": 768,
|
| 6 |
+
"num_hidden_layers": 12,
|
| 7 |
+
"intermediate_size": 2048,
|
| 8 |
+
"vocab_size": 50257,
|
| 9 |
+
"tie_word_embeddings": true,
|
| 10 |
+
"max_position_embeddings": 1024,
|
| 11 |
+
"norm_type": "rmsnorm",
|
| 12 |
+
"layer_norm_eps": 1e-12,
|
| 13 |
+
"rms_norm_eps": 1e-06,
|
| 14 |
+
"num_dense_layers": 0,
|
| 15 |
+
"num_attention_heads": 12,
|
| 16 |
+
"q_shared_dim": null,
|
| 17 |
+
"kv_shared_dim": 128,
|
| 18 |
+
"o_shared_dim": null,
|
| 19 |
+
"qk_private_dim": 64,
|
| 20 |
+
"vo_private_dim": 64,
|
| 21 |
+
"rope_dims": 32,
|
| 22 |
+
"nope_dims": 32,
|
| 23 |
+
"rope_theta": 10000.0,
|
| 24 |
+
"rope_scaling": {
|
| 25 |
+
"type": "linear",
|
| 26 |
+
"factor": 2.0
|
| 27 |
+
},
|
| 28 |
+
"attention_bias": false,
|
| 29 |
+
"attention_backend": "flash_attention_2",
|
| 30 |
+
"ffn_decompose": false,
|
| 31 |
+
"ffn_rank": null,
|
| 32 |
+
"vocab_subspace": false,
|
| 33 |
+
"vocab_rank": null,
|
| 34 |
+
"hidden_dropout_prob": 0.1,
|
| 35 |
+
"attention_dropout_prob": 0.1,
|
| 36 |
+
"classifier_dropout": null,
|
| 37 |
+
"initializer_range": 0.02
|
| 38 |
+
},
|
| 39 |
+
"pre_train": {
|
| 40 |
+
"wandb_project": "decoder-pretrain-c4",
|
| 41 |
+
"output_dir": "checkpoints/gpt-2_seq1024_mla0-128-0",
|
| 42 |
+
"seed": 42,
|
| 43 |
+
"logging_steps": 20,
|
| 44 |
+
"save_steps": 300,
|
| 45 |
+
"train_batch_size": 64,
|
| 46 |
+
"gradient_accumulation_steps": 16,
|
| 47 |
+
"learning_rate": 0.0005,
|
| 48 |
+
"num_train_steps": 3000,
|
| 49 |
+
"eval_steps": 300,
|
| 50 |
+
"weight_decay": 0.01,
|
| 51 |
+
"num_workers": 8,
|
| 52 |
+
"pin_memory": true,
|
| 53 |
+
"_comment_dataset": "Use preprocessed_dataset_path instead of dataset streaming",
|
| 54 |
+
"preprocessed_dataset_path": "/home/ubuntu/c4_en_pct0.02_seq1024/c4_en_pct0.02_seq1024/dataset",
|
| 55 |
+
"dataset_name": "allenai/c4",
|
| 56 |
+
"dataset_config": "en",
|
| 57 |
+
"dataset_subset_pct": 0.02,
|
| 58 |
+
"max_seq_length": 1024,
|
| 59 |
+
"eval_batch_size": 64,
|
| 60 |
+
"fp16": false,
|
| 61 |
+
"bf16": true,
|
| 62 |
+
"torch_compile": true,
|
| 63 |
+
"torch_compile_backend": "inductor",
|
| 64 |
+
"torch_compile_mode": "default",
|
| 65 |
+
"run_name": "107.76M - model.768.lyr.12 - seqlen.1024 - mla.0.128.0 - ah12.64 - rd32.32",
|
| 66 |
+
"run_id": "u647cues",
|
| 67 |
+
"run_url": "https://wandb.ai/chrismccormick/decoder-pretrain-c4/runs/u647cues",
|
| 68 |
+
"best_checkpoint": "checkpoints/gpt-2_seq1024_mla0-128-0/checkpoint-3000"
|
| 69 |
+
},
|
| 70 |
+
"stats": {
|
| 71 |
+
"total_elements": "107.76M"
|
| 72 |
+
}
|
| 73 |
+
}
|