feat: Use new llama_token_is_eog in create_completions

abetlen · abetlen · commit d40a250ef3cf · 2024-04-22T00:35:47.000-04:00
diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py
@@ -1034,7 +1034,8 @@ def logit_bias_processor(
             logits_processor=logits_processor,
             grammar=grammar,
         ):
-            if token == self._token_eos:
+            assert self._model.model is not None
+            if llama_cpp.llama_token_is_eog(self._model.model, token):
                 text = self.detokenize(completion_tokens, prev_tokens=prompt_tokens)
                 finish_reason = "stop"
                 break
