DeepSpeed

Unsloth

unslothai/unsloth

Trending

Fast, memory-efficient LoRA/QLoRA fine-tuning and RL for open LLMs on a single consumer GPU.

LlamaFactory

hiyouga/LlamaFactory

Popular

Unified fine-tuning toolkit with a web UI supporting 100+ LLMs and VLMs across SFT, DPO and more.

TRL

huggingface/trl

Popular

Hugging Face library for post-training LLMs with SFT, DPO, GRPO and other RL-based methods.

Axolotl

axolotl-ai-cloud/axolotl

Config-driven fine-tuning framework for LLMs supporting full, LoRA and QLoRA training at multi-GPU scale.

verl

verl-project/verl

Trending

ByteDance's flexible reinforcement learning library for LLM post-training (PPO, GRPO) at scale.

Megatron-LM

NVIDIA/Megatron-LM

NVIDIA's reference for training transformer models at massive scale with tensor and pipeline parallelism.

torchtitan

pytorch/torchtitan

PyTorch-native platform for large-scale LLM pretraining with FSDP2, tensor and pipeline parallelism.

bitsandbytes

bitsandbytes-foundation/bitsandbytes

8-bit and 4-bit quantization kernels that make QLoRA and low-memory inference possible.

Tinker

thinkingmachines.ai

New

Thinking Machines' managed API for fine-tuning open-weight models while you keep control of the training loop.