Unsloth

WeChat Official Account Article

Unsloth is a lightweight library designed to accelerate the fine-tuning of LLMs (Large Language Models). It is fully compatible with the Hugging Face ecosystem, including the Hub, transformers, and PEFT.

You can use Unsloth in conjunction with Transformers or TRL to accelerate LLM model training, while utilizing SwanLab for experiment tracking and visualization.

1. Introducing SwanLabCallback

from swanlab.integration.transformers import SwanLabCallback

SwanLabCallback is a logging class adapted for Transformers.

The parameters that can be defined for SwanLabCallback include:

• project, experiment_name, description, and other parameters consistent with swanlab.init, used for initializing the SwanLab project.
• You can also create a project externally via swanlab.init, and the integration will log the experiment to the project you created externally.

2. Passing to Trainer

from swanlab.integration.huggingface import SwanLabCallback
from trl import GRPOTrainer

...

# Instantiate SwanLabCallback
swanlab_callback = SwanLabCallback(project="unsloth-example")

trainer = GRPOTrainer(
    ...
    # Pass the callbacks parameter
    callbacks=[swanlab_callback],
)

trainer.train()

3. Example Template Combining Unsloth

from swanlab.integration.transformers import SwanLabCallback
from unsloth import FastLanguageModel, PatchFastRL

PatchFastRL("GRPO", FastLanguageModel)  # Patch TRL
from trl import GRPOConfig, GRPOTrainer, ModelConfig, TrlParser

...

model, tokenizer = FastLanguageModel.from_pretrained(
...
) 

# PEFT Model
model = FastLanguageModel.get_peft_model(
...
)

# Instantiate SwanLabCallback
swanlab_callback = SwanLabCallback(
  project="trl_integration",
  experiment_name="qwen2.5-sft",
  description="Testing the integration of swanlab and trl",
  config={"framework": "🤗TRL"},
)

# Define GRPOTrainer
trainer = GRPOTrainer(
    ...
    # Pass the callbacks parameter
    callbacks=[swanlab_callback],
)

# Start training!
trainer.train()