{
  "model_type": "llama",
  "hidden_size": 5120,                // Dimensionality of hidden layers
  "num_attention_heads": 40,          // Number of attention heads
  "num_hidden_layers": 40,            // Number of transformer layers
  "intermediate_size": 13824,         // Feedforward layer size (often ~2.7x hidden_size)
  "hidden_act": "silu",               // Activation function
  "rotary_emb_base": 10000,           // Base for rotary embeddings
  "max_position_embeddings": 2048,    // Maximum sequence length
  "initializer_range": 0.02,          // Weight initialization range
  "rms_norm_eps": 1e-6,               // Epsilon for RMSNorm
  "use_parallel_residual": true,      // Optimized residual connections
  "vocab_size": 32000,                // Vocabulary size for the tokenizer
  "pad_token_id": 0,                  // Padding token ID
  "eos_token_id": 2,                  // End-of-sequence token ID
  "bos_token_id": 1,                  // Beginning-of-sequence token ID
  "torch_dtype": "float16",           // Floating point precision for weights (fp16)
  "use_cache": true                   // Enable caching during inference
}