Table of contents

axolotl — LLM fine-tuning framework

OpenAccess-AI-Collective’s Axolotl — a YAML-driven framework for full fine-tunes, LoRA, QLoRA, and DPO across PyTorch, DeepSpeed, and FSDP backends.

Image tag

docker.io/manvarharsh/axolotl:cuda12

What’s in this image

  • Base: nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
  • Python 3.10 (conda)
  • PyTorch with CUDA 12
  • Axolotl + dependencies (transformers, accelerate, deepspeed, peft, bitsandbytes)
  • Flash Attention 2
  • JupyterHub with Podstack authenticator
  • OpenSSH server

Default ports

PortService
22SSH
8000JupyterHub

Use cases

  • Full fine-tunes of 7B–70B Llama / Mistral / Qwen models
  • QLoRA / LoRA on consumer GPUs
  • DPO / KTO preference tuning
  • Multi-GPU training via DeepSpeed / FSDP
  • Reproducible YAML-config training runs

Environment variables

VariableDescription
ENABLE_SSHEnable SSH server
ENABLE_JUPYTERHUBEnable JupyterHub on port 8000
PODSTACK_API_URLBackend URL for JupyterHub token validation
SSH_PUBLIC_KEYPublic key for SSH

Persistence

Mount at /data. Place training configs in /data/configs/, datasets in /data/datasets/, outputs in /data/output/.

See also