Merge pull request #75 from togethercomputer/add-streaming-tokens

Add streaming tokens

Author: csris

inference/bot.py

@@ -11,10 +11,39 @@
 import argparse
 import conversation as convo
 import retrieval.wikipedia as wp
-from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig
+from transformers import AutoTokenizer, AutoModelForCausalLM, AutoConfig, StoppingCriteria, StoppingCriteriaList
 from accelerate import infer_auto_device_map, init_empty_weights
 
 
+class StopWordsCriteria(StoppingCriteria):
+    def __init__(self, tokenizer, stop_words, stream_callback):
+        self._tokenizer = tokenizer
+        self._stop_words = stop_words
+        self._partial_result = ''
+        self._stream_buffer = ''
+        self._stream_callback = stream_callback
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        first = not self._partial_result
+        text = self._tokenizer.decode(input_ids[0, -1])
+        self._partial_result += text
+        for stop_word in self._stop_words:
+            if stop_word in self._partial_result:
+                return True
+        if self._stream_callback:
+            if first:
+                text = text.lstrip()
+            # buffer tokens if the partial result ends with a prefix of a stop word, e.g. "<hu"
+            for stop_word in self._stop_words:
+                for i in range(1, len(stop_word)):
+                    if self._partial_result.endswith(stop_word[0:i]):
+                        self._stream_buffer += text
+                        return False
+            self._stream_callback(self._stream_buffer + text)
+            self._stream_buffer = ''
+        return False
+
+
 class ChatModel:
     human_id = "<human>"
     bot_id = "<bot>"
@@ -54,7 +83,8 @@ def __init__(self, model_name, gpu_id, max_memory):
             )
         self._tokenizer = AutoTokenizer.from_pretrained(model_name)
 
-    def do_inference(self, prompt, max_new_tokens, do_sample, temperature, top_k):
+    def do_inference(self, prompt, max_new_tokens, do_sample, temperature, top_k, stream_callback=None):
+        stop_criteria = StopWordsCriteria(self._tokenizer, [self.human_id], stream_callback)
         inputs = (
             self._tokenizer(prompt, return_tensors='pt')
             .to(self._model.device)
@@ -65,7 +95,8 @@ def do_inference(self, prompt, max_new_tokens, do_sample, temperature, top_k):
             do_sample=do_sample,
             temperature=temperature,
             top_k=top_k,
-            pad_token_id=self._tokenizer.eos_token_id
+            pad_token_id=self._tokenizer.eos_token_id,
+            stopping_criteria=StoppingCriteriaList([stop_criteria]),
         )
         output = self._tokenizer.batch_decode(outputs)[0]
 
@@ -79,7 +110,7 @@ class OpenChatKitShell(cmd.Cmd):
     intro = "Welcome to OpenChatKit shell.   Type /help or /? to list commands.\n"
     prompt = ">>> "
 
-    def __init__(self, gpu_id, model_name_or_path, max_tokens, sample, temperature, top_k, retrieval, max_memory):
+    def __init__(self, gpu_id, model_name_or_path, max_tokens, sample, temperature, top_k, retrieval, max_memory, do_stream):
         super().__init__()
         self._gpu_id = int(gpu_id)
         self._model_name_or_path = model_name_or_path
@@ -89,6 +120,7 @@ def __init__(self, gpu_id, model_name_or_path, max_tokens, sample, temperature,
         self._top_k = top_k
         self._retrieval = retrieval
         self._max_memory = max_memory
+        self._do_stream = do_stream
 
     def preloop(self):
         print(f"Loading {self._model_name_or_path} to cuda:{self._gpu_id}...")
@@ -120,12 +152,13 @@ def do_say(self, arg):
             self._max_tokens,
             self._sample,
             self._temperature,
-            self._top_k
+            self._top_k,
+            lambda x : print(x, end='', flush=True) if self._do_stream else None,
         )
 
         self._convo.push_model_response(output)
 
-        print(self._convo.get_last_turn())
+        print("" if self._do_stream else self._convo.get_last_turn())
 
     def do_raw_say(self, arg):
         output = self._model.do_inference(
@@ -183,6 +216,11 @@ def main():
         action='store_true',
         help='indicates whether to sample'
     )
+    parser.add_argument(
+        '--no-stream',
+        action='store_true',
+        help='indicates whether to stream tokens'
+    )
     parser.add_argument(
         '--temperature',
         default=0.6,
@@ -238,7 +276,8 @@ def main():
         args.temperature,
         args.top_k,
         args.retrieval,
-        max_memory
+        max_memory,
+        not args.no_stream,
     ).cmdloop()