config.json

{
  "variant": "matformer",
  "vocab_size": 50432,
  "num_hidden_layers": 32,
  "hidden_size": 512,
  "head_dim": 128,
  "base_width": null,
  "mlp_base_width": 128,
  "widths": [
    4,
    9,
    15,
    22,
    30
  ],
  "tiers": [
    "low",
    "medium",
    "high",
    "extra",
    "max"
  ],
  "mlp_output_divisor": 0.25,
  "split_matrices": [
    "attn.q_proj",
    "attn.k_proj",
    "attn.v_proj",
    "attn.gate_proj",
    "attn.o_proj",
    "mlp.gate_proj",
    "mlp.up_proj",
    "mlp.down_proj"
  ],
  "rope_theta": 10000.0,
  "rope_precision": "bfloat16",
  "rms_norm_eps": 1e-05,
  "max_position_embeddings": 2048,
  "tie_word_embeddings": false,
  "tokenizer": "EleutherAI/gpt-neox-20b",
  "architectures": [
    "ElasticForCausalLM"
  ],
  "model_type": "elastic",
  "torch_dtype": "float32",
  "method": "MatFormer",
  "scale": "119M–285M"
}