Minimal change for latest transformer usage.

openicl/icl_dataset_reader.py

@@ -225,16 +225,21 @@ def __init__(self, datalist: List, model_name=None, tokenizer=None) -> None:
             self.tokenizer = tokenizer
         else:
             self.tokenizer = AutoTokenizer.from_pretrained(model_name)
-            self.tokenizer.pad_token = self.tokenizer.eos_token
-            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+            if self.tokenizer.eos_token is not None:
+                self.tokenizer.pad_token = self.tokenizer.eos_token
+            if self.tokenizer.eos_token_id is not None:
+                self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
             self.tokenizer.padding_side = "left"
         self.encode_dataset = []
         self.init_dataset()
         self.datalist_length = len(self.encode_dataset)
 
     def init_dataset(self):
         for idx, data in enumerate(self.datalist):
-            tokenized_data = self.tokenizer.encode_plus(data, truncation=True, return_tensors='pt', verbose=False)
+            try:
+                tokenized_data = self.tokenizer(data, truncation=True, return_tensors='pt', verbose=False)
+            except (TypeError, AttributeError):
+                tokenized_data = self.tokenizer.encode_plus(data, truncation=True, return_tensors='pt', verbose=False)
             self.encode_dataset.append({
                 'input_ids': tokenized_data.input_ids[0],
                 'attention_mask': tokenized_data.attention_mask[0],

openicl/icl_inferencer/icl_base_inferencer.py

@@ -141,8 +141,11 @@ def __init_tokenizer(self, tokenizer_name):
                 self.tokenizer = tokenizer_name
             else:
                 self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-        self.tokenizer.pad_token = self.tokenizer.eos_token
-        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+        if self.tokenizer.eos_token is not None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+        if self.tokenizer.eos_token_id is not None:
+            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
         self.tokenizer.padding_side = "left"
 
     def __init_api(self, **kwargs):

openicl/icl_inferencer/icl_ppl_inferencer.py

@@ -183,8 +183,9 @@ def __get_ppl(self, input_texts: List[str], mask_length=None):
                     mask[i][j] = 1
             loss = loss * mask
 
-        lens = (inputs["input_ids"] != self.tokenizer.pad_token_id).sum(-1).cpu().numpy()
+        lens = (inputs["input_ids"] != self.tokenizer.pad_token_id).sum(-1)
         if mask_length is not None:
-            lens -= np.array(mask_length)
-        ce_loss = loss.sum(-1).cpu().detach().numpy() / lens
+            lens -= torch.tensor(mask_length, device=lens.device, dtype=lens.dtype)
+        # Some new hf models are bfloat16
+        ce_loss = (loss.sum(-1) / lens.to(loss.dtype)).detach().to(torch.float32).cpu().numpy()
         return ce_loss

openicl/icl_retriever/icl_topk_retriever.py

@@ -62,8 +62,10 @@ def __init__(self,
         gen_datalist = self.dataset_reader.generate_input_field_corpus(self.test_ds)
 
         self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
-        self.tokenizer.pad_token = self.tokenizer.eos_token
-        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+        if self.tokenizer.eos_token is not None:
+            self.tokenizer.pad_token = self.tokenizer.eos_token
+        if self.tokenizer.eos_token_id is not None:
+            self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
         self.tokenizer.padding_side = "right"
 
         self.encode_dataset = DatasetEncoder(gen_datalist, tokenizer=self.tokenizer)

openicl/utils/collators.py

@@ -62,6 +62,8 @@ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) ->
         batch.update(res_dict)
 
         if self.device:
-            batch = batch.to(self.device)
+            for k, v in list(batch.items()):
+                if hasattr(v, "to"):
+                    batch[k] = v.to(self.device)
 
         return batch