allow custom stopping strings

[catalpaaa]

a better #823

[catalpaaa]

[oobabooga]

It now works in notebook mode too

[ShaneTRS]

I've just done git pull and updated all the requirements via pip, and I can't seem to verify that this is functioning. I'm using the api-example.py script, with custom_stopping_strings set to ["\n"] (as well as early_stopping set to True, though I'm not sure that's necessary). I also changed the prompt to "test", just because I know it often creates new lines.

I'm still getting responses that are multiple lines long, such as:

test_2.csv"
    train_df = pd.read_csv(train_csv)
    test_df = pd.read_csv(test_csv)
    # preprocessing

I've tried other strings (and other amounts of strings), such as first names and what-not, but none seem to have an effect.

[ShaneTRS]

After some extra testing, I think it may be related to the fact that I run a RWKV model.
Here are my typical launch arguments:
--model RWKV-4-Raven-3B-v8-EngAndMore-ctx4096.pth --rwkv-strategy 'cuda fp16' --rwkv-cuda-on --listen --no-stream --notebook

I tried running with these instead:
--model llama-7b --load-in-8bit --listen --no-stream --notebook --lora alpaca-lora-7b

I seem to get proper stopping, though I didn't do a lot of testing.

I'm not familiar with proper etiquette, so I haven't created a separate issue regarding this.

[oobabooga]

@ShaneTRS the stopping criteria stuff is a part of transformers, and RWKV uses a custom implementation that is not part of transformers, so it doesn't work for this model (and also for llama.cpp)

See https://github.com/oobabooga/text-generation-webui/blob/main/modules/text_generation.py#L139