[Feature] Add end_str for turbomind (#859)

* fix

* update

* fix internlm1

* fix docs

* remove sys

configs/eval_internlm_chat_turbomind.py

@@ -23,6 +23,14 @@ internlm_meta_template = dict(round=[
 ],
                               eos_token_id=103028)
 
+internlm2_meta_template = dict(
+    round=[
+        dict(role='HUMAN', begin='<|im_start|>user\n', end='<|im_end|>\n'),
+        dict(role='BOT', begin='<|im_start|>assistant\n', end='<|im_end|>\n', generate=True),
+    ],
+    eos_token_id=92542
+)
+
 # config for internlm-chat-7b
 internlm_chat_7b = dict(
     type=TurboMindModel,
@@ -41,6 +49,28 @@ internlm_chat_7b = dict(
     concurrency=32,
     meta_template=internlm_meta_template,
     run_cfg=dict(num_gpus=1, num_procs=1),
+    end_str='<eoa>',
+)
+
+# config for internlm-chat-7b
+internlm2_chat_7b = dict(
+    type=TurboMindModel,
+    abbr='internlm2-chat-7b-turbomind',
+    path='internlm/internlm2-chat-7b',
+    engine_config=dict(session_len=2048,
+                       max_batch_size=32,
+                       rope_scaling_factor=1.0),
+    gen_config=dict(top_k=1,
+                    top_p=0.8,
+                    temperature=1.0,
+                    max_new_tokens=100),
+    max_out_len=100,
+    max_seq_len=2048,
+    batch_size=32,
+    concurrency=32,
+    meta_template=internlm2_meta_template,
+    run_cfg=dict(num_gpus=1, num_procs=1),
+    end_str='<|im_end|>'
 )
 
 # config for internlm-chat-20b
@@ -61,6 +91,7 @@ internlm_chat_20b = dict(
     concurrency=8,
     meta_template=internlm_meta_template,
     run_cfg=dict(num_gpus=1, num_procs=1),
+    end_str='<eoa>',
 )
 
 models = [internlm_chat_20b]

docs/en/advanced_guides/evaluation_turbomind.md

@@ -56,6 +56,7 @@ internlm_20b = dict(
         batch_size=8,
         concurrency=8,
         run_cfg=dict(num_gpus=1, num_procs=1),
+        end_str='<eoa>'
     )
 
 models = [internlm_20b]

docs/zh_cn/advanced_guides/evaluation_turbomind.md

@@ -56,6 +56,7 @@ internlm_20b = dict(
         batch_size=8,
         concurrency=8,
         run_cfg=dict(num_gpus=1, num_procs=1),
+        end_str='<eoa>'
     )
 
 models = [internlm_20b]

opencompass/models/turbomind.py

@@ -34,6 +34,9 @@ class TurboMindModel(BaseModel):
             arguments like session_len, max_batch_size for TurboMind.
         gen_config (Dict, optional): Generation config to set
                 arguments like top_k, top_p, temperature.
+        end_str (str, optional): Whether to trim generated strings with end_str
+            if the model has special ending strings that are not handled well.
+            Defaults to None.
     """
 
     def __init__(self,
@@ -42,7 +45,8 @@ class TurboMindModel(BaseModel):
                  max_seq_len: int = 2048,
                  meta_template: Optional[Dict] = None,
                  engine_config: Optional[Dict] = None,
-                 gen_config: Optional[Dict] = None):
+                 gen_config: Optional[Dict] = None,
+                 end_str: Optional[str] = None):
         super().__init__(path=path,
                          max_seq_len=max_seq_len,
                          meta_template=meta_template)
@@ -62,6 +66,7 @@ class TurboMindModel(BaseModel):
         ]
         self.generator_ids = [i + 1 for i in range(concurrency)]
         self.gen_config = gen_config
+        self.end_str = end_str
 
     def generate(
         self,
@@ -90,11 +95,15 @@ class TurboMindModel(BaseModel):
         for batch_input in batch_inputs:
             with ThreadPoolExecutor() as executor:
                 _results = list(
-                    executor.map(self._generate,
+                    executor.map(
+                        self._generate,
                         self.generators[:len(batch_input)],
                         self.generator_ids[:len(batch_input)],
-                                 batch_input, [max_out_len] * len(batch_input),
-                                 [self.gen_config] * len(batch_input)))
+                        batch_input,
+                        [max_out_len] * len(batch_input),
+                        [self.gen_config] * len(batch_input),
+                        [self.end_str] * len(batch_input),
+                    ))
                 results += _results
         return results
 
@@ -114,7 +123,8 @@ class TurboMindModel(BaseModel):
                   session_id,
                   prompt: str or PromptList,
                   max_out_len: int,
-                  gen_config=None) -> str:
+                  gen_config=None,
+                  end_str: Optional[str] = None) -> str:
         """Generate results given a list of inputs.
 
         Args:
@@ -124,7 +134,10 @@ class TurboMindModel(BaseModel):
             max_out_len (int): The maximum length of the output.
             gen_config (EngineGenerationConfig, optional): Generation
                 config to set arguments like top_k, top_p, temperature.
-
+            end_str (str, optional): Whether to trim generated strings
+                with end_str if the model has special ending strings
+                that are not handled well.
+                Defaults to None.
         Returns:
             str: The generated string.
         """
@@ -144,4 +157,7 @@ class TurboMindModel(BaseModel):
             _, output_ids, _ = outputs
             response = self.tokenizer.decode(output_ids)
             response = valid_str(response)
+        # used to trim
+        if end_str:
+            response = response.split(end_str)[0]
         return response