Skip to content

feat: support Beam search ("Best of") setting #3112

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
1 task done
walking-octopus opened this issue Jun 27, 2024 · 1 comment
Closed
1 task done

feat: support Beam search ("Best of") setting #3112

walking-octopus opened this issue Jun 27, 2024 · 1 comment
Labels
P2: nice to have Nice to have feature type: feature request A new feature

Comments

@walking-octopus
Copy link

  • I have searched the existing issues

Is your feature request related to a problem? Please describe it

Beam search is a known technique used to enhance LLM inference, especially in tasks requiring reasoning, at a great cost to performance. ChatGPT has been suspected of using it.

Describe the solution

llama.cpp does support it. All that would be needed is to add somewhere in inference config.

Teachability, documentation, adoption, migration strategy

No response

What is the motivation / use case for changing the behavior?

Example: mathematical problems. Global probability maximization allows the model to catch and prune erroneous paths, since eventually they lead to less probable chains of thought than their incorrect counterparts.

@Van-QA Van-QA added the P2: nice to have Nice to have feature label Jun 30, 2024
@imtuyethan imtuyethan moved this to Icebox in Menlo Sep 2, 2024
@freelerobot freelerobot changed the title feat: Beam search ("Best of") setting feat: support Beam search ("Best of") setting Sep 5, 2024
@dan-menlo
Copy link
Contributor

@walking-octopus I am merging this into #3508 (Jan) and menloresearch/cortex.cpp#1151 (Cortex.cpp)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
P2: nice to have Nice to have feature type: feature request A new feature
Projects
Archived in project
Development

No branches or pull requests

4 participants