LLM Fine-Tuning Frameworks: Unsloth Leads Speed, Axolotl Scales

Four open-source projects—Unsloth, Axolotl, TRL, and LLaMA-Factory—are prominent for large language model (LLM) fine-tuning, all built upon the PyTorch and Hugging Face stack. They differentiate themselves through engineering focus: Unsloth optimizes by rewriting kernels, Axolotl composes parallelism strategies, TRL provides trainer APIs, and LLaMA-Factory prioritizes broad model coverage and zero-code operation. This comparison evaluates their performance across training throughput, peak VRAM usage, and multi-GPU scaling.

TRL serves as the foundational reference implementation, offering trainers like SFTTrainer, DPOTrainer, and GRPOTrainer, with its current stable release being v1.8.0. Both Axolotl and LLaMA-Factory integrate with TRL. Unsloth distinguishes itself by replacing parts of the modeling code with custom Triton kernels, deriving backpropagation steps manually rather than relying on autograd. Hugging Face's documentation indicates this approach results in 0% accuracy degradation compared to standard QLoRA, as no approximations are introduced.

Axolotl functions as a YAML-driven wrapper for Transformers, PEFT, TRL, Accelerate, and DeepSpeed, with its primary advantage being the composability of parallelism strategies rather than kernel-level optimizations. LLaMA-Factory, presented as an ACL 2024 system demonstration, features a Gradio web UI named LlamaBoard and supports over 100 LLMs and vision-language models (VLMs).

In terms of speed, Unsloth demonstrates significant gains on a single GPU. Its benchmarks show a 2x training speed increase for Llama 3.1 8B and Llama 3.3 70B using the Alpaca dataset, a batch size of 2, and gradient accumulation of 4, with QLoRA at rank 32. For larger models, Unsloth reported fine-tuning unsloth/gpt-oss-20b-BF16 on an NVIDIA B200, achieving 712.33 ms per step at an 8K context length, a substantial improvement over the 5,226 ms per step of a comparative setup.