diff --git a/llm_deepseek.py b/llm_deepseek.py index 45a71fa..3205f96 100644 --- a/llm_deepseek.py +++ b/llm_deepseek.py @@ -51,7 +51,7 @@ class Options(Chat.Options): def execute(self, prompt, stream, response, conversation): messages = self._build_messages(conversation, prompt) response._prompt_json = {"messages": messages} - kwargs = self.build_kwargs(prompt) + kwargs = self.build_kwargs(prompt, stream) max_tokens = kwargs.pop('max_tokens', 8192) if prompt.options.response_format: @@ -126,7 +126,7 @@ class Options(Completion.Options): def execute(self, prompt, stream, response, conversation): full_prompt = self._build_full_prompt(conversation, prompt) response._prompt_json = {"prompt": full_prompt} - kwargs = self.build_kwargs(prompt) + kwargs = self.build_kwargs(prompt, stream) max_tokens = kwargs.pop('max_tokens', 4096) if prompt.options.echo: