iamlemec
diff --git a/‎llama_cpp/llama.py
Copy file name to clipboardExpand all lines: llama_cpp/llama.py
+12-8Lines changed: 12 additions & 8 deletions b/‎llama_cpp/llama.py
Copy file name to clipboardExpand all lines: llama_cpp/llama.py
+12-8Lines changed: 12 additions & 8 deletions
@@ -948,8 +948,7 @@ def logit_bias_processor(
 
             if stream:
                 remaining_tokens = completion_tokens[returned_tokens:]
-                prev_tokens = completion_tokens[:returned_tokens]
-                remaining_text = self.detokenize(completion_tokens, prev_tokens)
+                remaining_text = self.detokenize(remaining_tokens)
                 remaining_length = len(remaining_text)
 
                 # We want to avoid yielding any characters from
@@ -971,13 +970,13 @@ def logit_bias_processor(
                     for token in remaining_tokens:
                         if token == self.token_bos():
                             continue
-                        token_end_position += len(remaining_text)
+                        token_end_position += len(self.detokenize([token]))
                         # Check if stop sequence is in the token
                         if token_end_position > (
                             remaining_length - first_stop_position
                         ):
                             break
-                        token_str = remaining_text.decode(
+                        token_str = self.detokenize([token]).decode(
                             "utf-8", errors="ignore"
                         )
                         text_offset = len(prompt) + len(
@@ -1002,7 +1001,11 @@ def logit_bias_processor(
                         }
                         top_logprob.update({token_str: current_logprobs[int(token)]})
                         logprobs_or_none = {
-                            "tokens": [token_str],
+                            "tokens": [
+                                self.detokenize([token]).decode(
+                                    "utf-8", errors="ignore"
+                                )
+                            ],
                             "text_offset": [text_offset],
                             "token_logprobs": [current_logprobs[int(token)]],
                             "top_logprobs": [top_logprob],
@@ -1015,7 +1018,9 @@ def logit_bias_processor(
                             "model": model_name,
                             "choices": [
                                 {
-                                    "text": token_str,
+                                    "text": self.detokenize([token]).decode(
+                                        "utf-8", errors="ignore"
+                                    ),
                                     "index": 0,
                                     "logprobs": logprobs_or_none,
                                     "finish_reason": None,
@@ -1027,7 +1032,7 @@ def logit_bias_processor(
                         decode_success = False
                         for i in range(1, len(remaining_tokens) + 1):
                             try:
-                                bs = remaining_text
+                                bs = self.detokenize(remaining_tokens[:i])
                                 ts = bs.decode("utf-8")
                                 decode_success = True
                                 break
@@ -1063,7 +1068,6 @@ def logit_bias_processor(
 
             if len(completion_tokens) >= max_tokens:
                 text = self.detokenize(completion_tokens)
-                
                 finish_reason = "length"
                 break