Method answer

demo/test.py:320–360

(self, conversations, language_model_inputs, modal_type="image")

    @torch.no_grad()
    def answer(self, conversations, language_model_inputs, modal_type="image"):
        model_inputs = self.tokenizer(
            conversations,
            return_tensors="pt",
        )
        model_inputs.pop("token_type_ids", None)

        input_ids = model_inputs["input_ids"].to(self.device)
        attention_mask = model_inputs["attention_mask"].to(self.device)

        if modal_type == "text":
            generation_output = self.model.language_model.generate(
                input_ids=input_ids,
                attention_mask=attention_mask,
                generation_config=self.generation_config,
                return_dict_in_generate=True,
                output_scores=True
            )
        else:
            pixel_values = model_inputs.pop("pixel_values", None)
            if pixel_values is not None:
                pixel_values = pixel_values.to(self.device)

            generation_output = self.model.generate(
                pixel_values=pixel_values,
                input_ids=input_ids,
                attention_mask=attention_mask,
                language_model_inputs=language_model_inputs,
                generation_config=self.generation_config,
                return_dict_in_generate=True,
                output_scores=True
            )

        preds = generation_output.sequences
        outputs = self.tokenizer.batch_decode(preds, skip_special_tokens=True)[0]

        if modal_type == "text":
            skip_echo_len = len(conversations[0]) - conversations[0].count("</s>") * 3
            outputs = outputs[skip_echo_len:].strip()

        return outputs

if __name__ == '__main__':
    model_path = “your path”