{
  "architectures": [
    "QMoEForCausalLM"
  ],
  "auto_map": {
    "AutoConfig": "configuration_qmoe.QMoEConfig",
    "AutoModel": "modeling_qmoe.QMoEForCausalLM",
    "AutoModelForCausalLM": "modeling_qmoe.QMoEForCausalLM"
  },
  "bos_token_id": 50256,
  "d_model": 768,
  "dtype": "float32",
  "eos_token_id": 50256,
  "ffn_dim": 2048,
  "is_decoder": true,
  "max_seq_len": 512,
  "model_type": "qmoe",
  "moe_top_k": 2,
  "num_experts": 8,
  "num_heads": 16,
  "num_layers": 12,
  "pad_token_id": 50256,
  "tie_word_embeddings": false,
  "transformers_version": "4.57.3",
  "use_cache": false,
  "vocab_size": 50257
}