config.json

{
  "model_type": "ChatGLM2Model",
  "model_config": {
    "hidden_size": 4096,
    "inner_hidden_size": 13696,
    "head_hidden_size": 128,
    "num_multi_query_groups": 2,
    "num_attention_heads": 32,
    "num_layers": 28,
    "vocab_size": 65024,
    "dropout_rate": 0.0,
    "layernorm_epsilon": 1e-05,
    "max_sequence_length": 8192
  },
  "quant_type": "int8",
  "weight_files": [
    "model_weights_0.safetensors",
    "model_weights_1.safetensors",
    "model_weights_2.safetensors"
  ],
  "tokenizer_file": "sentencepiece.model",
  "torch_dtype": "float16"
}