full_config.json

{
  "shorthand": "model.768.lyr.12 - seqlen.1024 - mla.0.0.0 - ah12.64 - rd32.32",
  "notes": "GPT-2 run with essentially MHA.",
  "model": {
    "hidden_size": 768,
    "num_hidden_layers": 12,
    "intermediate_size": 2048,
    "vocab_size": 50257,
    "tie_word_embeddings": true,
    "max_position_embeddings": 1024,
    "norm_type": "rmsnorm",
    "layer_norm_eps": 1e-12,
    "rms_norm_eps": 1e-06,
    "num_dense_layers": 0,
    "num_attention_heads": 12,
    "q_shared_dim": null,
    "kv_shared_dim": null,
    "o_shared_dim": null,
    "qk_private_dim": 64,
    "vo_private_dim": 64,
    "rope_dims": 32,
    "nope_dims": 32,
    "rope_theta": 10000.0,
    "rope_scaling": {
      "type": "linear",
      "factor": 2.0
    },
    "attention_bias": false,
    "attention_backend": "flash_attention_2",
    "ffn_decompose": false,
    "ffn_rank": null,
    "vocab_subspace": false,
    "vocab_rank": null,
    "hidden_dropout_prob": 0.1,
    "attention_dropout_prob": 0.1,
    "classifier_dropout": null,
    "initializer_range": 0.02
  },
  "pre_train": {
    "wandb_project": "decoder-pretrain-c4",
    "output_dir": "checkpoints/gpt-2_seq1024_mla0-0-0",
    "seed": 42,
    "logging_steps": 20,
    "save_steps": 300,
    "train_batch_size": 64,
    "gradient_accumulation_steps": 16,
    "learning_rate": 0.0005,
    "num_train_steps": 3000,
    "eval_steps": 300,
    "weight_decay": 0.01,
    "num_workers": 8,
    "pin_memory": true,
    "_comment_dataset": "Use preprocessed_dataset_path instead of dataset streaming",
    "preprocessed_dataset_path": "/home/ubuntu/c4_en_pct0.02_seq1024/c4_en_pct0.02_seq1024/dataset",
    "dataset_name": "allenai/c4",
    "dataset_config": "en",
    "dataset_subset_pct": 0.02,
    "max_seq_length": 1024,
    "eval_batch_size": 64,
    "fp16": false,
    "bf16": true,
    "torch_compile": true,
    "torch_compile_backend": "inductor",
    "torch_compile_mode": "default",
    "run_name": "117.88M - model.768.lyr.12 - seqlen.1024 - mla.0.0.0 - ah12.64 - rd32.32",
    "run_id": "15ixey47",
    "run_url": "https://wandb.ai/chrismccormick/decoder-pretrain-c4/runs/15ixey47",
    "best_checkpoint": "checkpoints/gpt-2_seq1024_mla0-0-0/checkpoint-3000"
  },
  "stats": {
    "total_elements": "117.88M"
  }
}