Environment:Huggingface Trl Python Core Dependencies

Knowledge Sources	TRL pyproject.toml
Domains	Infrastructure, Deep_Learning
Last Updated	2026-02-06 17:00 GMT

Overview

Python 3.10+ environment with transformers >= 4.56.2, accelerate >= 1.4.0, datasets >= 3.0.0, and PyTorch as the core runtime for all TRL trainers.

Description

This environment defines the mandatory runtime dependencies for the TRL library. Every trainer (SFT, DPO, GRPO, Reward, PPO, RLOO, and all experimental trainers) requires these packages at their specified minimum versions. The library is built on top of Hugging Face's transformers ecosystem and uses accelerate for distributed training orchestration and datasets for data loading.

Usage

Use this environment for any TRL training workflow. This is the base prerequisite that must be satisfied before any optional dependencies (vLLM, PEFT, DeepSpeed, quantization) can be layered on top.

System Requirements

Category	Requirement	Notes
OS	Linux (Ubuntu 20.04+), macOS, Windows (WSL2)	Linux is the primary supported platform
Hardware	CPU or GPU	GPU (NVIDIA CUDA, AMD ROCm, Intel XPU) recommended for training
Python	>= 3.10	Supports 3.10, 3.11, 3.12, 3.13
Disk	2GB+ for packages	Additional space needed for model weights and datasets

Dependencies

System Packages

`python` >= 3.10
`pip` (latest)

Python Packages

`transformers` >= 4.56.2
`accelerate` >= 1.4.0
`datasets` >= 3.0.0
`packaging` > 20.0
`torch` (PyTorch, version compatible with transformers)
`numpy`
`pandas`

Credentials

The following environment variables may be needed depending on usage:

`HF_TOKEN`: Hugging Face API token for accessing gated models and datasets, and for pushing models to the Hub.
`WANDB_API_KEY`: Weights & Biases API key for experiment tracking (optional).

Quick Install

# Install TRL with core dependencies
pip install trl

# Or install from source
pip install git+https://github.com/huggingface/trl.git

Code Evidence

Core dependencies from `pyproject.toml:30-35`:

requires-python = ">=3.10"
dependencies = [
    "accelerate>=1.4.0",
    "datasets>=3.0.0",
    "packaging>20.0",
    "transformers>=4.56.2",
]

Python version classifiers from `pyproject.toml:24-28`:

"Programming Language :: Python :: 3.10",
"Programming Language :: Python :: 3.11",
"Programming Language :: Python :: 3.12",
"Programming Language :: Python :: 3.13"

Hardware detection from `trl/scripts/env.py:42-47`:

if torch.cuda.is_available():
    devices = [torch.cuda.get_device_name(i) for i in range(torch.cuda.device_count())]
elif torch.backends.mps.is_available():
    devices = ["MPS"]
elif torch.xpu.is_available():
    devices = [torch.xpu.get_device_name(i) for i in range(torch.xpu.device_count())]

Common Errors

Error Message	Cause	Solution
`requires-python >= 3.10`	Python version too old	Upgrade to Python 3.10+
`ImportError: transformers`	transformers not installed or version too old	`pip install transformers>=4.56.2`
`ImportError: accelerate`	accelerate not installed	`pip install accelerate>=1.4.0`

Compatibility Notes

NVIDIA GPUs (CUDA): Full support. Primary development platform.
AMD GPUs (ROCm): Supported via PyTorch ROCm builds.
Intel XPU: Supported via PyTorch XPU builds. Detected via torch.xpu.is_available().
Apple MPS: Detected but limited support for training.
CPU: Supported for development and testing; not recommended for training.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment