better use of tokenizer

Signed-off-by: n1ck-guo <[email protected]>

auto_round/mllm/autoround_mllm.py

@@ -41,12 +41,13 @@ def _only_text_test(model, tokenizer, device, model_type):
     if model_type in SUPPORT_ONLY_TEXT_MODELS:  # save time
         return True
 
+    new_tokenizer = deepcopy(tokenizer)
     device = detect_device(device)
     text = ["only text", "test"]
-    tokenizer.padding_side = 'left'
-    if tokenizer.pad_token is None:
-        tokenizer.pad_token = tokenizer.eos_token
-    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
+    new_tokenizer.padding_side = 'left'
+    if new_tokenizer.pad_token is None:
+        new_tokenizer.pad_token = new_tokenizer.eos_token
+    inputs = new_tokenizer(text, return_tensors="pt", padding=True, truncation=True)
 
     try:
         inputs = inputs.to(device)
@@ -182,7 +183,7 @@ def __init__(
         if isinstance(dataset, str):
             if quant_nontext_module or \
                 (dataset in CALIB_DATASETS.keys() and not \
-                 _only_text_test(model, deepcopy(tokenizer), device, self.template.model_type)):
+                 _only_text_test(model, tokenizer, device, self.template.model_type)):
                 if quant_nontext_module:
                     logger.warning(f"Text only dataset cannot be used for calibrating non-text modules,"
                                 "switching to liuhaotian/llava_conv_58k")