Fixed error for replace -100s in predictions by the pad token (intel#353)

Signed-off-by: Cheng, Penghui <[email protected]>

Author: PenghuiCheng

examples/huggingface/pytorch/summarization/quantization/run_summarization.py

@@ -651,6 +651,8 @@ def compute_metrics(eval_preds):
         preds, labels = eval_preds
         if isinstance(preds, tuple):
             preds = preds[0]
+        # Replace -100s used for padding as we can't decode them
+        preds = np.where(preds != -100, preds, tokenizer.pad_token_id)
         decoded_preds = tokenizer.batch_decode(preds, skip_special_tokens=True)
         if data_args.ignore_pad_token_for_loss:
             # Replace -100 in the labels as we can't decode them.

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -69,7 +69,7 @@ def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
         if load_in_8bit or load_in_4bit or quantization_config is not None:
             from intel_extension_for_transformers.llm.quantization.utils import convert_to_quantized_model
             torch_dtype = kwargs.pop("torch_dtype", torch.float32)
-        
+
         if load_in_4bit:
             if quantization_config is None:
                 quantization_config = WeightOnlyQuantConfig(compute_dtype=torch_dtype, weight_dtype="nf4")
@@ -204,6 +204,7 @@ def default_calib_func(model):
             )
         return model
 
+
 class AutoModelForCausalLM(_BaseQBitsAutoModelClass):
     ORIG_MODEL = transformers.AutoModelForCausalLM
 
@@ -214,4 +215,3 @@ class AutoModel(_BaseQBitsAutoModelClass):
 
 class AutoModelForSeq2SeqLM(_BaseQBitsAutoModelClass):
     ORIG_MODEL = transformers.AutoModelForSeq2SeqLM
-