config.json

{
  "model_type": "moe_transformer",
  "architectures": [
    "TransformerLanguageModel"
  ],
  "vocab_size": 16384,
  "d_model": 256,
  "num_heads": 8,
  "head_dim": 32,
  "num_layers": 4,
  "d_ff": 1024,
  "max_seq_len": 256,
  "num_experts": 4,
  "top_k": 2,
  "shared_routed": false,
  "matched_active": true,
  "aux_loss_coef": 0.01,
  "torch_dtype": "float32",
  "transformers_version": "4.40.0"
}