intel / xFasterTransformer

Apache License 2.0
355 stars 61 forks source link

qwen1.5-32b long text input issue #411

Closed zhm-algo closed 3 months ago

pujiang2018 commented 4 months ago

could you pls give more details with examples?

zhm-algo commented 4 months ago

在对 32B 进行了错误边缘检测后发现,当模型的输入文本大小大于 1.5K 以上就会出现,生成异常的问题。 如果输入文本大小在 1.5k-4k 这个区间内, 会不断地循环输出错误内容。 如果输入文本大小大于 4K 时, 模型响应结果是不断地输出回车符。

输入使用prompt.json 文件中 qwen 对应的8192/4096,可以复现问题 https://github.com/intel/xFasterTransformer/blob/main/benchmark/prompt.json

pujiang2018 commented 4 months ago

@marvin-Yu How about this issue, any finding?

marvin-Yu commented 4 months ago

The latest version does not reproduce the issue, @zhm-algo you can try the latest version again.

pujiang2018 commented 3 months ago

Let's close since 2 weeks passed. @zhm-algo pls reopen it if the issue is still there.