diff --git a/basaran/model.py b/basaran/model.py index c1bda2d7..470a449d 100644 --- a/basaran/model.py +++ b/basaran/model.py @@ -78,7 +78,7 @@ def __call__( temperature=temperature, top_p=top_p, ), - **kwargs + **kwargs, } # Generate completion tokens. @@ -347,4 +347,4 @@ def load_model( if not model.can_generate(): raise TypeError(f"{name_or_path} is not a text generation model") - return StreamModel(model, tokenizer) + return StreamModel(model.eval(), tokenizer)