config.json

{
  "_attn_implementation_autoset": true,
  "_name_or_path": "/data-fsx/swiftkv-llama33-70b-pretrain-math-1p5m-prod",
  "architectures": [
    "LlamaSwiftKVForCausalLM"
  ],
  "attention_bias": false,
  "attention_dropout": 0,
  "bos_token_id": 128000,
  "eos_token_id": [
    128001,
    128008,
    128009
  ],
  "head_dim": 128,
  "hidden_act": "silu",
  "hidden_size": 8192,
  "initializer_range": 0.02,
  "intermediate_size": 28672,
  "key_value_group_size": 1,
  "max_position_embeddings": 131072,
  "mlp_bias": false,
  "model_type": "llama_swiftkv",
  "num_attention_heads": 64,
  "num_hidden_layers": 80,
  "num_key_value_heads": 8,
  "num_key_value_layers": 40,
  "pretraining_tp": 1,
  "rms_norm_eps": 1e-05,
  "rope_scaling": {
    "factor": 8,
    "high_freq_factor": 4,
    "low_freq_factor": 1,
    "original_max_position_embeddings": 8192,
    "rope_type": "llama3"
  },
  "rope_theta": 500000,
  "swiftkv": true,
  "tie_word_embeddings": false,
  "torch_dtype": "bfloat16",
  "transformers_version": "4.47.0",
  "use_cache": true,
  "vocab_size": 128256
}