Add repetition_penalty aligned with huggingface

[Abraham-Xu]

In huggingface/transformers generate method:
https://github.com/huggingface/transformers/blob/ae320fa53f74cc4dfa0e4fc3c95b6129a86b0512/src/transformers/generation/utils.py#L1295

1. it prepares a logits_processor from generation_config
   https://github.com/huggingface/transformers/blob/ae320fa53f74cc4dfa0e4fc3c95b6129a86b0512/src/transformers/generation/utils.py#L1540
2. it does logits pre-process by logits_processor in greedy_search/sample/beam_search
   https://github.com/huggingface/transformers/blob/ae320fa53f74cc4dfa0e4fc3c95b6129a86b0512/src/transformers/generation/utils.py#L2457

Specificaly, repetition penalty is a frequently used pre-process method in logits_processor. It prevents repetitions through a penalty. This penalized sampling works by discounting the scores of previously generated tokens. (https://arxiv.org/pdf/1909.05858.pdf)

I introduced the the repetition penalty pre-process to sampler.py aligned with huggingface implementation:
https://github.com/huggingface/transformers/blob/ae320fa53f74cc4dfa0e4fc3c95b6129a86b0512/src/transformers/generation/logits_process.py#L328

[Scagin]

I am waiting for this, I need some method to make hf and vllm compat when I use penalty arguments

[Abraham-Xu]

I have tested the repetition_penalty under greedy search mode by Llama 7b and Baichuan models. The results are all aligned with huggingface transformers.

[markluofd]

this pr solved my problem, for my llama2 model, repetition_penalty is necessary, otherwise the result is incorrect

[TheBloke]

Great! Looking forward to testing this as my perception is that vLLM is very poor with repetition currently, and that the existing repetition penalty params (freqency_penalty etc) do little or nothing.

Although part of that problem is that there's no per-request seed, something we also really need.

[jeffchy]

Beautiful！

[BaiMoHan]

Very nice

[BaiMoHan]

Hey, what's wrong?I did not find the repetition_penalty attribute in sampling_params.py when I use v0.1.7. I'm a bit impatient.🥺🥺🥺

[leshanbog]

this comment is misleading now?

[leshanbog]

why else? will presence_penalty and repetition_penalty work together?

[heshuguo]

@WoosukKwon please check if the fix can be merged.

[WoosukKwon]

@Abraham-Xu @heshuguo This PR is a bit in conflict with #1048, which tries to refactor the sampler code. @zhuohan123 Can we merge this PR before yours?

[WoosukKwon]

Hi @Abraham-Xu, sorry for the late response.

Could you update the PR? There are some merge conflicts because we refactored sampler in #1048.

[jessiewiswjc]

I want to know if presence_penalty/frequency_penalty in vllm can work with repetition_penalty in hf at the same time?

[zhuohan123]

Supported with #1424. But again, thank you for your contribution! Let us know if there is any other issue.

[Abraham-Xu]

Supported with #1424. But again, thank you for your contribution! Let us know if there is any other issue.

Sorry for late reply, busy with working and other things recently. It is great to support repetition penalty in any way.