[Good First Issue]: causal_lm/cpp must read EOS token value from rt_info of openvino_tokenizer.xml

[p-wysocki]

Context

End Of Sequence tokens are an essential part of LLM training and inference. You can find more details in this comment.

Thanks to a PR adding End Of Sequence tokens to Runtime Info openvino_tokenizers now put EOS token value into rt_info section in OpenVINO Intermediate Representation format (.xml file to be specific) when converting a tokenizer to OpenVINO.

Since EOS has been enabled in OpenVINO, now it needs to be enabled in GenAI text_generation module.

What needs to be done?

beam_search_causal_lm.cpp and greedy_causal_lm.cpp from https://github.com/openvinotoolkit/openvino.genai/tree/master/text_generation/causal_lm/cpp should read the EOS token instead of having a hardcoded value with comment // There's no way to extract special token values from the detokenizer for now.

It’s required to extract the value using ov::Model::get_rt_info() and use it. Remove the comments about absence of way to extract that value.

Example Pull Requests

Resources

• Contribution guide - start here!
• Intel DevHub Discord channel - engage in discussions, ask questions and talk to OpenVINO developers

Contact points

@pavel-esir

Ticket

132861

[anzr299]

.take

[p-wysocki]

Hello @anzr299, the ticket will be manually assigned to you shortly - the take feature hasn't been introduced to GenAI repo yet, but it will be today. :)

Thanks for taking a look at the issue!

[p-wysocki]

Hello @anzr299, are you still working on this? Is there anything we could help you with?

[anzr299]

Hi, I am still working on it, was busy the past few days.

[anzr299]

I've created a PR #315

[anzr299]

I have a small question, in beam_search_casual_lm.cpp, the following lines check for an empty token and not the hardcoded EOS token. Should I change this as well to check for the EOS token?
if (next_tokens.empty()) { break; }
(lines 60-62)

[p-wysocki]

cc @pavel-esir

[pavel-esir]

I have a small question, in beam_search_casual_lm.cpp, the following lines check for an empty token and not the hardcoded EOS token. Should I change this as well to check for the EOS token? if (next_tokens.empty()) { break; } (lines 60-62)

good question! No need to update beam_searcher, just initialize it with eos token you got from IR in the sample. Please look my comments in you PR.

[anzr299]

@ilya-lavrenov The issue asked to add a functionality to read EOS tokens only for beam_search_causal_lm and greedy_causal_lm but not for speculative_decoding_lm. I followed the requirement in my work but I thought it could be better if I added the functionality to speculative_decoding_lm too. Is that alright?

[ilya-lavrenov]

I followed the requirement in my work but I thought it could be better if I added the functionality to speculative_decoding_lm too. Is that alright?

Yes, let's handle this sample as well.
The initial requirements were added before the speculative_decoding_lm sample is implemented.

[anzr299]

I followed the requirement in my work but I thought it could be better if I added the functionality to speculative_decoding_lm too. Is that alright?

Yes, let's handle this sample as well. The initial requirements were added before the speculative_decoding_lm sample is implemented.

Sure, I will create a new PR

[anzr299]

@ilya-lavrenov I've created a PR at #353