Update code for slim

config.json

@@ -8,9 +8,11 @@
     "AutoModel": "modeling_chatglm.ChatGLMForConditionalGeneration",
     "AutoModelForSeq2SeqLM": "modeling_chatglm.ChatGLMForConditionalGeneration"
   },
-  "bos_token_id": 150004,
-  "eos_token_id": 150005,
-  "pad_token_id": 20003,
+  "bos_token_id": 130004,
+  "eos_token_id": 130005,
+  "mask_token_id": 130000,
+  "gmask_token_id": 130001,
+  "pad_token_id": 3,
   "hidden_size": 4096,
   "inner_hidden_size": 16384,
   "layernorm_epsilon": 1e-05,
@@ -22,5 +24,5 @@
   "torch_dtype": "float16",
   "transformers_version": "4.23.1",
   "use_cache": true,
-  "vocab_size": 150528
+  "vocab_size": 130528
 }

configuration_chatglm.py

@@ -66,6 +66,8 @@ class ChatGLMConfig(PretrainedConfig):
             use_cache=False,
             bos_token_id=150004,
             eos_token_id=150005,
+            mask_token_id=150000,
+            gmask_token_id=150001,
             pad_token_id=0,
             max_sequence_length=2048,
             inner_hidden_size=16384,
@@ -86,6 +88,8 @@ class ChatGLMConfig(PretrainedConfig):
         self.bos_token_id = bos_token_id
         self.eos_token_id = eos_token_id
         self.pad_token_id = pad_token_id
+        self.mask_token_id = mask_token_id
+        self.gmask_token_id = gmask_token_id
         self.position_encoding_2d = position_encoding_2d
         self.quantization_bit = quantization_bit
         self.pre_seq_len = pre_seq_len

modeling_chatglm.py

@@ -921,7 +921,7 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
 
 
             if position_ids is None:
-                MASK, gMASK = 150000, 150001
+                MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
                 mask_token = MASK if MASK in input_ids else gMASK
                 use_gmask = False if MASK in input_ids else True
 
@@ -1084,7 +1084,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
             **kwargs
     ) -> dict:
         batch_size, seq_length = input_ids.shape
-        MASK, gMASK = 150000, 150001
+        MASK, gMASK = self.config.mask_token_id, self.config.gmask_token_id
         mask_token = MASK if MASK in input_ids else gMASK
         use_gmask = False if MASK in input_ids else True
         seqs = input_ids.tolist()

tokenization_chatglm.py

@@ -48,11 +48,13 @@ class SPTokenizer:
     def __init__(
             self,
             vocab_file,
+            num_image_tokens=20000,
             max_blank_length=80,
             byte_fallback=True,
     ):
         assert vocab_file is not None
         self.vocab_file = vocab_file
+        self.num_image_tokens = num_image_tokens
         self.special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "<unused_0>", "<sop>", "<eop>", "<ENC>", "<dBLOCK>"]
         self.max_blank_length = max_blank_length
         self.byte_fallback = byte_fallback
@@ -70,10 +72,6 @@ class SPTokenizer:
     def get_tab_token():
         return f"<|tab|>"
 
-    @property
-    def num_image_tokens(self):
-        return 20000
-
     @property
     def num_text_tokens(self):
         return self.text_tokenizer.num_tokens
@@ -178,6 +176,7 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
             mask_token='[MASK]',
             gmask_token='[gMASK]',
             padding_side="left",
+            num_image_tokens=20000,
             **kwargs
     ) -> None:
         super().__init__(
@@ -197,7 +196,7 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
         self.mask_token = mask_token
         self.gmask_token = gmask_token
 
-        self.sp_tokenizer = SPTokenizer(vocab_file)
+        self.sp_tokenizer = SPTokenizer(vocab_file, num_image_tokens=num_image_tokens)
 
         """ Initialisation """
 

tokenizer_config.json

@@ -10,6 +10,7 @@
   "remove_space": false,
   "do_lower_case": false,
   "tokenizer_class": "ChatGLMTokenizer",
+  "num_image_tokens": 0,
   "auto_map": {
     "AutoTokenizer": [
       "tokenization_chatglm.ChatGLMTokenizer",