token间没有space的情况

[chuyi369]

请问我在使用Yi-1.5-9B-Chat进行推理时，出现token间没有space的情况，这是bug吗？

transformers=4.47

            outputs = self.model(input_ids=input_ids, attention_mask=attention_mask, use_cache=True)
            next_token_logits = outputs.logits[:, -1, :] / temp
            next_token_logits = next_token_logits.to(self.model.device) 
            next_token_logits = self.filter_forbidden_tokens(next_token_logits)
            next_token_logits = logits_processor(input_ids, next_token_logits)
            output_tokens += 1
            step += 1

            next_token_id = torch.multinomial(F.softmax(next_token_logits, dim=-1), num_samples=1)

            next_word = self.tokenizer.decode(next_token_id[0], skip_special_tokens=True)

            response += next_word

[chuyi369]

谢谢我解决了这个问题，通过修改tokenizer_config.json中add_prefix_space为false

参考：#36