Create config.json

config.json

@@ -0,0 +1,25 @@
+{
+  "model_type": "llama",
+  "hidden_size": 4096,  // Adjust based on LLaMA size (8B uses 4096)
+  "intermediate_size": 11008,  // Typically 2.67x the hidden size
+  "num_attention_heads": 32,  // Adjust based on model
+  "num_hidden_layers": 32,  // Number of transformer layers
+  "vocab_size": 32000,  // Common LLaMA vocab size
+  "max_position_embeddings": 2048,  // Max sequence length you used
+  "rope_scaling": {
+    "type": "dynamic",
+    "factor": 2.0
+  },
+  "rotary_embedding_base": 10000,  // Used for RoPE embeddings
+  "tie_word_embeddings": false,  // Generally false for LLaMA
+  "torch_dtype": "float16",  // Use float16 for efficient inference
+  "use_cache": true,  // Enable cache for inference
+  "pad_token_id": 0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "quantization_config": {
+    "load_in_4bit": true,
+    "bnb_4bit_use_double_quant": true,
+    "bnb_4bit_quant_type": "nf4"
+  }
+}