config.json

{
  "architectures": [
    "Step1MoEForCausalLM"
  ],
  "model_type": "step1",
  "hidden_size": 1024,
  "intermediate_size": 9552,
  "num_attention_heads": 16,
  "num_attention_groups": 16,
  "num_hidden_layers": 8,
  "max_seq_len": 65536,
  "vocab_size": 65536,
  "rms_norm_eps": 1e-05,
  "torch_dtype": "bfloat16",
  "moe_every_n_layer": 64,
  "moe_intermediate_size": 4096,
  "moe_num_experts": 8,
  "moe_top_k": 2,
  "use_moe": true,
  "moe_layer_offset": 1
}