Axolotl
App in the BluixApps catalog
What it is
Axolotl is a config-driven LLM fine-tuning toolkit by OpenAccess AI Collective — supports LoRA/QLoRA/full-parameter training, multi-GPU via DeepSpeed/FSDP, and broad model coverage (Llama, Mistral, Qwen, Gemma, ChatGLM, Phi, etc.). The industry standard for production LLM fine-tuning.
When AI startups fine-tune their own LLMs, Axolotl is the most common choice.
What it's for
- LoRA / QLoRA fine-tuning — adapter training with low VRAM
- Full-parameter SFT — supervised fine-tuning end-to-end
- DPO / ORPO / KTO — preference alignment / RLHF alternatives
- Continued pretraining — extend base model on new domain
- Multi-GPU training — DeepSpeed ZeRO 1/2/3, FSDP
- Dataset prep — alpaca, sharegpt, jsonl formats supported
Who it's for
- AI startups fine-tuning custom LLMs for their domain
- Research teams publishing fine-tuned model variants
- Enterprises training internal-data-aware models
- Compliance-conscious teams keeping training data on-prem
- Hosting providers offering managed fine-tuning to clients
Why teams pick Axolotl over alternatives
- Apache 2.0 — fully open
- Config-driven — YAML files define entire training run
- Broad model support — every major HF base model works
- DeepSpeed integration — multi-GPU production-grade
- Active maintenance — OpenAccess collective + contributors
- Community recipes — example configs for popular use cases
- Used by major fine-tunes — Hermes, OpenHermes, etc. published with Axolotl
Integrations
- HuggingFace Transformers — base library
- PEFT — LoRA / QLoRA / IA3 / Prefix tuning
- TRL — SFTTrainer, DPOTrainer, ORPOTrainer
- bitsandbytes — 4/8-bit quantization
- DeepSpeed — ZeRO optimization for multi-GPU
- WandB / TensorBoard — training monitoring
- Pair with: vLLM/TGI to serve fine-tuned model post-training
Notable users & community
- 9k+ GitHub stars
- OpenAccess AI Collective backing
- Used to train NousResearch Hermes, Teknium models, OpenHermes series
- Many published HF fine-tunes credit Axolotl in their model cards
- Active Discord with researchers + practitioners
Tips & operations
- VRAM budgets:
- 7B QLoRA: 16 GB
- 7B LoRA: 24 GB
- 13B QLoRA: 24 GB
- 70B QLoRA: 80+ GB (or 2× 80 GB)
- Dataset format: JSONL with
{instruction, input, output}(alpaca) or{conversations: [...]}(sharegpt) - Config-driven: start with
examples/configs, modify - Multi-GPU:
accelerate launch --num-processes N -m axolotl.cli.train config.yml - DeepSpeed: enable ZeRO 2 or 3 for 70B+ models
- Monitoring: WandB integration for loss curves
- Output: LoRA adapter file + merged weights option
What we ship in BluixApps
- Docker (axolotlai/axolotl:main-latest)
- JupyterLab pre-installed for interactive training
- Persistent volumes: workspace, datasets, outputs
- Port 8888 mapped (Jupyter lab interface)
- Pre-set HF_TOKEN environment variable for gated models
- Install report at
/root/bluixapps/axolotl.txt - Quick-start commands for LoRA training
- Multi-GPU launch example
- Pairing notes (vLLM/TGI for serving fine-tuned)
- GPU pre-flight check via
bluixapps_ensure_nvidia_runtime - Backup hook covers workspace + outputs (datasets opt-in)
Get this app — pick a BluixApps plan
Same catalog. Scaling tenant isolation, white-label and support tier.
| Tier | Tenants | Catalog | Support | White-label | Monthly | |
|---|---|---|---|---|---|---|
| Stacks | 1 | 19 curated stacks | Standard | — | $19/mo | DetailDeploy |
| Starter | 10 | Full catalog | Standard | +$15–25/mo | $49/mo | DetailDeploy |
| Pro | 25 | Full catalog | Priority bugfix | +$15–25/mo | $149/mo | DetailDeploy |
| Growth | 100 | Full catalog | Priority bugfix | +$15–25/mo | $349/mo | DetailDeploy |
| Scale | 500 | Full catalog | 7-day window | +$15–25/mo | $799/mo | DetailDeploy |
| Enterprise | Unlimited | Full catalog | Priority 7-day | Bundled | $1,499/mo | DetailDeploy |
