pytorch-nlp-setup

name	pytorch-nlp-setup
description	Set up Python environments for NLP/ML projects using PyTorch, transformers, and TRL. Use this skill when installing dependencies for preference optimization, RLHF, or transformer-based training pipelines.

PyTorch NLP Environment Setup

Standard Stack for Preference Optimization Projects

pip install torch transformers datasets accelerate trl peft wandb numpy

Version Compatibility Notes

• TRL (Transformer Reinforcement Learning) provides base trainers like CPOTrainer, DPOTrainer
• Older TRL versions (< 0.8) have different import paths for utilities
• Check from trl.trainer.utils import DPODataCollatorWithPadding availability
• For trl_sanitze_kwargs_for_tagging (note: typo is intentional in some versions)

Common Issues

• If from trl.import_utils import is_peft_available fails, check TRL version
• Some projects pin specific transformers/TRL versions — check requirements.txt or setup.py
• CUDA availability: use torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

Running Unit Tests

cd /path/to/project && python -m pytest unit_test/ -v
# or
cd /path/to/project && python -m unittest unit_test.unit_test_1