[WebUI] Add prompt format and stopping words for Qwen (#10066)

* add prompt format and stopping_words for qwen mdoel

* performance optimization

* optimize

* update

* meet comments

python/llm/example/Text-Generation-WebUI/modules/callbacks.py

@@ -41,6 +41,18 @@ class _StopEverythingStoppingCriteria(transformers.StoppingCriteria):
         return shared.stop_everything
 
 
+class StopWordsCriteria(transformers.StoppingCriteria):
+    """Custom `StoppingCriteria` which checks if all generated functions in the batch are completed."""
+    def __init__(self, stop_words, tokenizer):
+        self.stop_words = stop_words
+        self.tokenizer = tokenizer
+
+    def __call__(self, input_ids, scores, **kwargs):
+        """Returns true if all generated sequences contain any of the end-of-function strings."""
+        text =  self.tokenizer.decode(input_ids[-1][-1])
+        return text in self.stop_words
+
+
 class Stream(transformers.StoppingCriteria):
     def __init__(self, callback_func=None):
         self.callback_func = callback_func

python/llm/example/Text-Generation-WebUI/modules/text_generation.py

@@ -35,7 +35,8 @@ import modules.shared as shared
 from modules.callbacks import (
     Iteratorize,
     Stream,
-    _StopEverythingStoppingCriteria
+    _StopEverythingStoppingCriteria,
+    StopWordsCriteria
 )
 from modules.extensions import apply_extensions
 from modules.grammar.grammar_utils import initialize_grammar
@@ -331,6 +332,19 @@ def generate_reply_HF(question, original_question, seed, state, stopping_strings
     if shared.args.deepspeed:
         generate_params.update({'synced_gpus': True})
 
+    #tune the prompt based on qwen
+    QWEN_PROMPT_FORMAT = """
+    <|im_start|>system
+    You are a helpful assistant.
+    <|im_end|>
+    <|im_start|>user
+    {prompt}
+    <|im_end|>
+    <|im_start|>assistant
+    """
+    if shared.model.config.model_type == "qwen":
+        question = QWEN_PROMPT_FORMAT.format(prompt=question)
+
     # Encode the input
     input_ids = encode(question, add_bos_token=state['add_bos_token'], truncation_length=get_max_prompt_length(state))
     output = input_ids[0]
@@ -346,10 +360,19 @@ def generate_reply_HF(question, original_question, seed, state, stopping_strings
         generate_params.update({'inputs_embeds': inputs_embeds})
 
     # Stopping criteria / eos token
+    generate_params['stopping_criteria'] = transformers.StoppingCriteriaList()
     eos_token_ids = [shared.tokenizer.eos_token_id] if shared.tokenizer.eos_token_id is not None else []
     generate_params['eos_token_id'] = eos_token_ids
-    generate_params['stopping_criteria'] = transformers.StoppingCriteriaList()
-    generate_params['stopping_criteria'].append(_StopEverythingStoppingCriteria())
+
+    if shared.model.config.model_type == "qwen":
+        stopping_words = ["<|endoftext|>", "<|im_end|>", "<|im_start|>"]
+        generate_params['stopping_criteria'].append(StopWordsCriteria(stopping_words, shared.tokenizer))
+
+    for st in state['custom_stopping_strings']:
+        if type(st) is str:
+            stopping_words = [item.strip().strip('"') for item in [state['custom_stopping_strings']][0].split(',')]
+            generate_params['stopping_criteria'].append(StopWordsCriteria(stopping_words, shared.tokenizer))
+
 
     # Logits processor
     processor = state.get('logits_processor', LogitsProcessorList([]))