bug fix

config.json

@@ -153,7 +153,7 @@
   },
   "max_position_embeddings": 32768,
   "mlp_bias": false,
-  "model_type": "llama",
+  "model_type": "mistral",
   "num_attention_heads": 32,
   "num_hidden_layers": 32,
   "num_key_value_heads": 8,

maker.sh

@@ -8,11 +8,11 @@ then TMPA=./maker$$a.py
 src="Rakuten/RakutenAI-7B"
 tgt="exRakutenAI-7B"
 import json,torch,unicodedata
-from transformers import LlamaTokenizerFast,LlamaForCausalLM
+from transformers import LlamaTokenizerFast,MistralForCausalLM
 tkz=LlamaTokenizerFast.from_pretrained(src,cls_token="<s>",sep_token="<s>",mask_token="<unk>",pad_token="</s>")
 d=json.loads(tkz.backend_tokenizer.to_str())
 tkz.backend_tokenizer.from_str(json.dumps(d)).save("tokenizer.json")
-mdl=LlamaForCausalLM.from_pretrained(src)
+mdl=MistralForCausalLM.from_pretrained(src)
 tkz=LlamaTokenizerFast(tokenizer_file="tokenizer.json",model_max_length=mdl.config.max_position_embeddings,cls_token="<s>",sep_token="<s>",mask_token="<unk>",pad_token="</s>")
 e=mdl.resize_token_embeddings(len(tkz))
 f=mdl.get_output_embeddings()