[Feature] 支持frequency_penalty采样
Open
backlog
- Dominant language
- Python
- Stars
- 8.1k
- Forks
- 748
- Avg merge
- 6d 2h
- Merged PRs (30d)
- 54
Description
### Motivation
当前lmdeploy支持repetition_penalty和presence_penalty。
而实际上只有repetition_penalty有效，LlamaBatch的sampling_params_仅初始化了以下几个参数列表：
可以参考vllm支持frequency_penalty采样吗，frequency_penalty与presence_penalty规则类似，区别在于，presence_penalty只对出现过的token减去一次penalty，而frequency_penalty会对出现过的token减去n次penalty（n为token在序列中出现的次数）。也就是说相同的token出现次数越多，影响越大。
### Related resources
_No response_
### Additional context
_No response_
Contributor guide
Assessment
This issue has not been assessed yet.