Implementation:Axolotl ai cloud Axolotl Setup Script

Knowledge Sources	Axolotl
Domains	Build_System, Packaging
Last Updated	2026-02-07 00:00 GMT

Overview

Legacy setuptools script that dynamically resolves Axolotl dependencies based on the detected PyTorch version and platform, managing compatibility across torch 2.4-2.9+ with appropriate xformers and vllm versions.

Description

The setup.py implements Axolotl's complex dependency resolution logic in the parse_requirements function. It reads requirements.txt, detects the installed PyTorch version, and dynamically selects compatible versions of xformers, vllm, fbgemm-gpu, and other CUDA-dependent packages. The resolution matrix covers PyTorch versions from 2.4 through 2.9+ with specific version pins for each minor/patch combination. On macOS, it strips incompatible packages (bitsandbytes, triton, mamba-ssm, xformers, liger-kernel). The extras_require map defines optional dependency groups: flash-attn, ring-flash-attn, deepspeed, mamba-ssm, auto-gptq, mlflow, galore, apollo, optimizers, ray, vllm, llmcompressor, fbgemm-gpu, and opentelemetry. The get_package_version function reads the version from the VERSION file.

Usage

This script is called by pip/setuptools during package installation. It works alongside pyproject.toml to handle the dynamic dependency resolution that cannot be expressed in static TOML configuration. It should be maintained when new PyTorch versions are released to ensure correct dependency pins.

Code Reference

Source Location

• Repository: Axolotl
• File: setup.py
• Lines: 1-211

Signature

def parse_requirements(extras_require_map: dict) -> tuple[list, list, dict]:
    """Parse requirements.txt and resolve version-specific dependencies.

    Args:
        extras_require_map: Mutable dict of optional dependency groups.

    Returns:
        Tuple of (install_requires, dependency_links, extras_require).
    """

def get_package_version() -> str:
    """Read package version from VERSION file."""

# Module-level execution:
extras_require = { "flash-attn": [...], "deepspeed": [...], ... }
install_requires, dependency_links, extras_require_build = parse_requirements(extras_require)
setup(
    version=get_package_version(),
    package_dir={"": "src"},
    packages=find_packages("src"),
    install_requires=install_requires,
    extras_require=extras_require_build,
)

Import

# Not imported directly. Executed by pip:
pip install -e .
# Or:
python setup.py develop

I/O Contract

Inputs

Name	Type	Required	Description
requirements.txt	File	Yes	Base package requirements
VERSION	File	Yes	Package version string
Installed torch version	Runtime detection	No	Auto-detected; falls back to 2.8.0
Platform	Runtime detection	No	macOS detection for skipping incompatible packages

Outputs

Name	Type	Description
install_requires	list[str]	Resolved core dependencies with version pins
dependency_links	list[str]	PyPI index URLs for CUDA wheels
extras_require	dict[str, list]	Optional dependency groups with version-resolved pins

Usage Examples

Standard Installation

# Install with PyTorch auto-detection
pip install -e .

# Install with specific extras
pip install -e ".[flash-attn,deepspeed]"

# Install with vllm support (requires compatible torch)
pip install -e ".[vllm]"

Related Pages

• Environment:Axolotl_ai_cloud_Axolotl_Python_Runtime