config.json

{
  "model_name": "emma-4",
  "model_type": "gpt",
  "architecture": "decoder-only-transformer-gqa",
  "vocab_size": 50000,
  "hidden_size": 1024,
  "num_hidden_layers": 24,
  "num_attention_heads": 16,
  "num_kv_heads": 4,
  "intermediate_size": 2752,
  "context_length": 2048,
  "rope_theta": 10000.0,
  "norm_type": "rmsnorm",
  "activation": "swiglu",
  "tokenizer": "sentencepiece",
  "tied_embeddings": true,
  "alignment": "SFT",
  "gpu_memory_limit_gb": 14.0,
  "onnx_opset": 18
}