NVIDIA-NeMo / NVIDIA-NeMo/Export-Deploy

Refactor tokenizer handling in nemo.export module

Open
#411 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

enhancement
Dominant language
Python
Stars
42
Forks
18
Avg merge
1d 2h
Merged PRs (30d)
8

Description

Is your feature request related to a problem? Please describe.
The logic for for instantiating tokenizers in TensorRTLLM class in nemo.export.tensorrt_llm has become complex:

There is NeMo 1 and NeMo 2 checkpoint support

There are two paths: using full NeMo setup (with io.load_context) or local implementations/copies in sentencepiece_tokenizer.py and tiktoken_tokenizer.py

There are different types of tokenizers: HF, binary tokenizer.model, Tiktoken
The task is to simplify this logic (one idea is to extend all the classes above with save_pretrained method and unify constructors if possible).

There is also no way of testing a Tiktoken-based model as there is no such checkpoint in CI system. We should add it.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start in nemo.export.tensorrt_llm and trace how TensorRTLLM instantiates NeMo 1, NeMo 2, Hugging Face, binary, and Tiktoken tokenizers. Read sentencepiece_tokenizer.py and tiktoken_tokenizer.py, then inspect the CI checkpoint setup. Done means the tokenizer paths are simpler and CI exercises a Tiktoken-based model.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
machine-learning, tooling
Issue type
Refactor
Difficulty
5/5
Estimated time
Over a week
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.