Correcting from Llama to Mistral

Correcting from Llama to Mistral saves memory when training/inference

[Check](https://huggingface.co/Vikhrmodels/Vikhr-7b-0.1/discussions/2)

config.json

@@ -1,7 +1,7 @@
 {
   "_name_or_path": "Vikhrmodels/Vikhr-7b-0.1",
   "architectures": [
-    "LlamaForCausalLM"
+    "MistralForCausalLM"
   ],
   "attention_bias": false,
   "attention_dropout": 0.0,
@@ -12,7 +12,7 @@
   "initializer_range": 0.02,
   "intermediate_size": 14336,
   "max_position_embeddings": 32768,
-  "model_type": "llama",
+  "model_type": "mistral",
   "num_attention_heads": 32,
   "num_hidden_layers": 32,
   "num_key_value_heads": 8,