unslothai/unsloth

[Bug] After adding import unsloth to the first line of the script, the GRPOTrainer fails to run properly; however, it works normally again once this import is removed. The Sophia optimizer interface being used was generated by an AI.

Open

#3591 opened on Nov 12, 2025

View on GitHub
 (11 comments) (0 reactions) (0 assignees)Python (64,271 stars) (5,658 forks)batch import
help wanted

Description

sophia_grpo.py Scripts generated by AI may encounter many issues, such as the inability to utilize multiple GPUs using fp16

xl.py token_utils.py

Contributor guide

Tech stack
pythonpytorch
Domain
machine learning
Issue type
bug
DifficultyEstimated implementation difficulty for a new contributor, from 1 for very small changes to 5 for expert-level work.
4
Estimated timeA rough time range for an experienced contributor to investigate, implement, test, and prepare a pull request.
half day
Activity statusHow available the issue appears right now: fresh, active, stale, blocked, or waiting on maintainer input.
active
ClarityHow clearly the issue explains the expected change, acceptance criteria, and next step.
unclear
Prerequisites
Python programmingPyTorch basicsUnsloth library
Newbie friendlinessA 1-100 score estimating how approachable this issue is for first-time contributors.
15
Research direction
First, examine the attached scripts (sophia grpo.py, xl.py, token utils.py) to understand the setup. The issue suggests that importing unsloth at the top of the script conflicts with GRPOTrainer. Look for any import order dependencies or shared module variables. Check the unsloth source code, particularly the GRPOTrainer class, for interactions with the Sophia optimizer. Review the 11 comments for any debugging insights from the community or maintainers.