support dynamic_eos (modelscope#1947)

swift/llm/utils/argument.py

@@ -135,7 +135,8 @@ def check_flash_attn(self: Union['SftArguments', 'InferArguments']) -> None:
     def handle_generation_config(self: Union['SftArguments', 'InferArguments']) -> None:
         if self.temperature == 0:
             self.do_sample = False
-        if self.do_sample is False:
+        if self.do_sample is False and (isinstance(self, InferArguments) and self.infer_backend == 'pt'
+                                        and isinstance(self, SftArguments)):
             # fix warning
             self.temperature = 1.
             self.top_p = 1.
@@ -994,7 +995,6 @@ def __post_init__(self) -> None:
             self.dataset_seed = self.seed
         self.set_model_type()
         self.check_flash_attn()
-        self.handle_generation_config()
         self.handle_lr_scheduler_kwargs()
         self.is_multimodal = self._is_multimodal(self.model_type)
         self.is_vision = self._is_vision(self.model_type)
@@ -1170,6 +1170,7 @@ def __post_init__(self) -> None:
             self.logging_dir = f'{self.output_dir}/runs'
             if self.train_backend == 'transformers':
                 self.training_args.logging_dir = self.logging_dir
+        self.handle_generation_config()
 
     def _init_training_args(self) -> None:
         additional_saved_files = []
@@ -1460,7 +1461,6 @@ def __post_init__(self) -> None:
         self.handle_custom_dataset_info()
         self.set_model_type()
         self.check_flash_attn()
-        self.handle_generation_config()
         self.is_multimodal = self._is_multimodal(self.model_type)
         self.prepare_ms_hub()
 
@@ -1492,6 +1492,7 @@ def __post_init__(self) -> None:
             self.sft_type = 'full'
 
         self.handle_infer_backend()
+        self.handle_generation_config()
 
     def handle_infer_backend(self):
         model_info = MODEL_MAPPING[self.model_type]

swift/llm/utils/template.py

@@ -806,6 +806,19 @@ def _encode_context_list(
             loss_scale.extend([loss_weight] * len(token_list))
         return input_ids, labels, loss_scale, tokenizer_kwargs
 
+    @staticmethod
+    def use_dynamic_eos(labels: List[int], suffix_tokens_id: List[int]) -> None:
+        suffix_len = len(suffix_tokens_id)
+        start = 0
+        for i in range(1, len(labels)):
+            if labels[i - 1] >= 0 and labels[i] == -100:
+                start = i
+            if start > 0 and labels[i - 1] == -100 and labels[i] >= 0:
+                # [0, 1, 2, -100(start), -100, 3(i), 4]
+                length = i - start
+                if length >= suffix_len:
+                    labels[start:start + suffix_len] = suffix_tokens_id
+
     def _concat_and_tokenize(self,
                              query: str,
                              query_role: str,
@@ -840,18 +853,10 @@ def _concat_and_tokenize(self,
         history.append([query, response])
         history_roles.append([query_role, 'assistant'])
 
-        # Set the loss_scale of chat_sep or suffix to 1 if efficient_eos.
-        efficient_eos = False
-        if self.chat_sep is not None and len(self.chat_sep) > 0:
-            if isinstance(self.chat_sep[0], str) and isinstance(self.suffix[0], str) and self.chat_sep[0].startswith(
-                    self.suffix[0]):
-                efficient_eos = True
-            elif isinstance(self.chat_sep[0], list) and self.chat_sep[0] == self.suffix[0]:
-                efficient_eos = True
-
         for i, ((q, r), (qr, rr)) in enumerate(zip(history, history_roles)):
             context_list = self.tool_prompt.copy() if qr == 'tool' else prompt.copy()
             extra_context_list = []
+            is_suffix = False
             if i < len(history) - 1:
                 context_list = [context for context in context_list if '{{SYSTEM}}' not in context]
                 context_list.append('{{RESPONSE}}')
@@ -861,14 +866,16 @@ def _concat_and_tokenize(self,
                 # last response
                 context_list.append('{{RESPONSE}}')
                 extra_context_list = self.suffix
-                efficient_eos = True
+                is_suffix = True
             if q or r:
                 self._concat_context_list(
                     context_list, res_context_list, loss_scale_list, query=q, response=r, system=system, round0=i)
                 res_context_list += extra_context_list
-                loss_scale_list += ([1.] if efficient_eos else [0.]) * len(extra_context_list)
+                loss_scale_list += ([1.] if is_suffix else [0.]) * len(extra_context_list)
         res_context_list, loss_scale_list = self._simplify_context_list(res_context_list, loss_scale_list, **kwargs)
         input_ids, labels, loss_scale, tokenizer_kwargs = self._encode_context_list(res_context_list, loss_scale_list)
+        if labels is not None:
+            self.use_dynamic_eos(labels, self._encode_context_list(self.suffix)[0])
 
         if response is None:
             labels = None

swift/llm/utils/utils.py

@@ -382,8 +382,8 @@ def _is_special(token: int) -> bool:
         if not _is_special(input_ids[i]) and _is_special(input_ids[i - 1]):
             e = i
             result_str += f'[{input_ids[i - 1]} * {e - s}]'
-    if _is_special(input_ids[-1]):
-        result_str += f'[{input_ids[i - 1]} * {len(input_ids) - s}]'
+    if _is_special(input_ids[i]):
+        result_str += f'[{input_ids[i]} * {len(input_ids) - s}]'
     else:
         result_str += tokenizer.decode(input_ids[e:], **tokenizer_kwargs)
     return result_str