I'm unable to replicate the experimental results presented in the paper.

[gigigigig1]

I used the Dolly-15k dataset, Llama-2-7B-chat model and an A800 device to make some experiments between your lookahead and lookahead-decoding, but I'm unable to replicate the results on the table2 of the paper. Even your lookahead was not as good as the lookahead-decoding.

[gigigigig1]

请问论文实验结果中的基线方法的测量结果（即不采用任何加速方法）是如何得来的，我将你们测试代码中lookahead为false的情况与直接调用transformers库中的generate方法测试的速度不一样，因此我不知道如何对比你们的lookahead方法和lookahead_decoding方法的效果

[zheyishine]

Our lookahead must be fully warmed up with thousands of requests to fill the trie tree. Using half of the samples to warmup would reproduction the result.