ggml-org/llama.cpp

feature request - disabling tokenizer in conversion / inference

Closed

#1,765 opened on Jun 8, 2023

View on GitHubContributor guide
 (7 comments) (1 reaction) (0 assignees)C++ (21,737 forks)batch import
enhancementgood first issuehelp wanted

Repository metrics

Stars
 (124,043 stars)
PR merge metrics
 (Avg merge 6d 8h) (389 merged PRs in 30d)

Description

in #1764 i asked if it'd be possible to add a Huggingface tokenizer. but - HF tokenizers are quite flexible and officially supporting them in llama.cpp (or ggml?) might be a lot of hassle.

a much easier workaround would be allowing to disable tokenizers in both model conversion and inference. this means the users are supposed to encode(text)/decode(ids) in their implementation for using llama.cpp. in my case, for example, i'll use a python GUI and a wrapper anyway.

i'd like to work on it, but honestly i don't think i understand enough to be able to do this. i'd appreciate very much if anyone's interested in it.

Contributor guide

Research direction
Examine the tokenizer integration in llama.cpp by reviewing the conversion script and inference code. Refer to the discussion in linked issue #1764 for context. Identify where tokenizer encoding/decoding occurs and consider adding a configuration flag to bypass it.
Tech stack
ccpp
Domain
backendmachine learning
Issue type
Feature
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
3
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
1-2 days
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
Stale
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
Mostly clear
Prerequisites
Basic C++Understanding of tokenizer internals
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
45