RangiLyu commited on
Commit
fd2aa27
1 Parent(s): a68e778

update chat template in model

Browse files
Files changed (1) hide show
  1. modeling_internlm2.py +6 -6
modeling_internlm2.py CHANGED
@@ -1138,12 +1138,12 @@ class InternLM2ForCausalLM(InternLM2PreTrainedModel):
1138
  def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = [], meta_instruction=""):
1139
  prompt = ""
1140
  if meta_instruction:
1141
- prompt += f"""<s>[UNUSED_TOKEN_146]system\n{meta_instruction}[UNUSED_TOKEN_145]\n"""
1142
  else:
1143
  prompt += "<s>"
1144
  for record in history:
1145
- prompt += f"""[UNUSED_TOKEN_146]user\n{record[0]}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n{record[1]}[UNUSED_TOKEN_145]\n"""
1146
- prompt += f"""[UNUSED_TOKEN_146]user\n{query}[UNUSED_TOKEN_145]\n[UNUSED_TOKEN_146]assistant\n"""
1147
  return tokenizer([prompt], return_tensors="pt")
1148
 
1149
  @torch.no_grad()
@@ -1165,7 +1165,7 @@ class InternLM2ForCausalLM(InternLM2PreTrainedModel):
1165
  inputs = self.build_inputs(tokenizer, query, history, meta_instruction)
1166
  inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
1167
  # also add end-of-assistant token in eos token id to avoid unnecessary generation
1168
- eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids(["[UNUSED_TOKEN_145]"])[0]]
1169
  outputs = self.generate(
1170
  **inputs,
1171
  streamer=streamer,
@@ -1178,7 +1178,7 @@ class InternLM2ForCausalLM(InternLM2PreTrainedModel):
1178
  )
1179
  outputs = outputs[0].cpu().tolist()[len(inputs["input_ids"][0]) :]
1180
  response = tokenizer.decode(outputs, skip_special_tokens=True)
1181
- response = response.split("[UNUSED_TOKEN_145]")[0]
1182
  history = history + [(query, response)]
1183
  return response, history
1184
 
@@ -1231,7 +1231,7 @@ class InternLM2ForCausalLM(InternLM2PreTrainedModel):
1231
  return
1232
 
1233
  token = self.tokenizer.decode([value[-1]], skip_special_tokens=True)
1234
- if token.strip() != "[UNUSED_TOKEN_145]":
1235
  self.response = self.response + token
1236
  history = self.history + [(self.query, self.response)]
1237
  self.queue.put((self.response, history))
 
1138
  def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = [], meta_instruction=""):
1139
  prompt = ""
1140
  if meta_instruction:
1141
+ prompt += f"""<s><|im_start|>system\n{meta_instruction}<|im_end|>\n"""
1142
  else:
1143
  prompt += "<s>"
1144
  for record in history:
1145
+ prompt += f"""<|im_start|>user\n{record[0]}<|im_end|>\n<|im_start|>assistant\n{record[1]}<|im_end|>\n"""
1146
+ prompt += f"""<|im_start|>user\n{query}<|im_end|>\n<|im_start|>assistant\n"""
1147
  return tokenizer([prompt], return_tensors="pt")
1148
 
1149
  @torch.no_grad()
 
1165
  inputs = self.build_inputs(tokenizer, query, history, meta_instruction)
1166
  inputs = {k: v.to(self.device) for k, v in inputs.items() if torch.is_tensor(v)}
1167
  # also add end-of-assistant token in eos token id to avoid unnecessary generation
1168
+ eos_token_id = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids(["<|im_end|>"])[0]]
1169
  outputs = self.generate(
1170
  **inputs,
1171
  streamer=streamer,
 
1178
  )
1179
  outputs = outputs[0].cpu().tolist()[len(inputs["input_ids"][0]) :]
1180
  response = tokenizer.decode(outputs, skip_special_tokens=True)
1181
+ response = response.split("<|im_end|>")[0]
1182
  history = history + [(query, response)]
1183
  return response, history
1184
 
 
1231
  return
1232
 
1233
  token = self.tokenizer.decode([value[-1]], skip_special_tokens=True)
1234
+ if token.strip() != "<|im_end|>":
1235
  self.response = self.response + token
1236
  history = self.history + [(self.query, self.response)]
1237
  self.queue.put((self.response, history))