Tokenizer overwritten

unsloth/models/llama.py

@@ -916,6 +916,7 @@ class FastLlamaModel:
         rope_scaling   = None,
         fix_tokenizer  = True,
         model_patcher  = None,
+        tokenizer_name = None,
         **kwargs,
     ):
         if model_patcher is None: model_patcher = FastLlamaModel
@@ -978,18 +979,17 @@ class FastLlamaModel:
             max_position_embeddings = max_position_embeddings,
             **kwargs,
         )
+
+        # Counteract saved tokenizers
+        tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
         tokenizer = AutoTokenizer.from_pretrained(
-            model_name,
+            tokenizer_name,
             model_max_length = max_position_embeddings,
             padding_side     = "right",
             token            = token,
         )
-
+        
-        print(tokenizer)
-        print(tokenizer.chat_template)
         model, tokenizer = patch_tokenizer(model, tokenizer)
-        print(tokenizer)
-        print(tokenizer.chat_template)
         model = model_patcher.post_patch(model)
 
         # Patch up QKV / O and MLP

unsloth/models/loader.py

@@ -18,7 +18,7 @@ from transformers import AutoConfig
 from transformers import __version__ as transformers_version
 from peft import PeftConfig, PeftModel
 from .mapper import INT_TO_FLOAT_MAPPER, FLOAT_TO_INT_MAPPER
-
+import os
 
 # https://github.com/huggingface/transformers/pull/26037 allows 4 bit loading!
 major, minor = transformers_version.split(".")[:2]
@@ -79,14 +79,12 @@ class FastLanguageModel(FastLlamaModel):
     ):
         old_model_name = model_name
         model_name = _get_model_name(model_name, load_in_4bit)
-        print(model_name)
 
         # First check if it's a normal model via AutoConfig
         is_peft = False
         try:
             model_config = AutoConfig.from_pretrained(model_name, token = token)
             is_peft = False
-            print(model_config)
         except:
             try:
                 # Most likely a PEFT model
@@ -98,7 +96,6 @@ class FastLanguageModel(FastLlamaModel):
             model_name = _get_model_name(peft_config.base_model_name_or_path, load_in_4bit)
             model_config = AutoConfig.from_pretrained(model_name, token = token)
             is_peft = True
-            print(model_config)
         pass
 
         model_type = model_config.model_type
@@ -121,7 +118,16 @@ class FastLanguageModel(FastLlamaModel):
             )
         pass
 
-        print(model_name)
+        # Check if this is local model since the tokenizer gets overwritten
+        if  os.path.exists(os.path.join(old_model_name, "tokenizer_config.json")) and \
+            os.path.exists(os.path.join(old_model_name, "tokenizer.json")) and \
+            os.path.exists(os.path.join(old_model_name, "special_tokens_map.json")):
+
+            tokenizer_name = old_model_name
+        else:
+            tokenizer_name = None
+        pass
+
         model, tokenizer = dispatch_model.from_pretrained(
             model_name     = model_name,
             max_seq_length = max_seq_length,
@@ -132,6 +138,7 @@ class FastLanguageModel(FastLlamaModel):
             rope_scaling   = rope_scaling,
             fix_tokenizer  = fix_tokenizer,
             model_patcher  = dispatch_model,
+            tokenizer_name = tokenizer_name,
             *args, **kwargs,
         )
 

unsloth/models/mistral.py

@@ -294,6 +294,7 @@ class FastMistralModel(FastLlamaModel):
         rope_scaling   = None, # Mistral does not support RoPE scaling
         fix_tokenizer  = True,
         model_patcher  = None,
+        tokenizer_name = None,
         **kwargs,
     ):
         if model_patcher is None: model_patcher = FastMistralModel
@@ -354,8 +355,11 @@ class FastMistralModel(FastLlamaModel):
             # rope_scaling      = rope_scaling,
             **kwargs,
         )
+
+        # Counteract saved tokenizers
+        tokenizer_name = model_name if tokenizer_name is None else tokenizer_name
         tokenizer = AutoTokenizer.from_pretrained(
-            model_name,
+            tokenizer_name,
             model_max_length = max_position_embeddings,
             padding_side     = "right",
             token            = token,