训练框架#

• deepspeedai/DeepSpeed
• unslothai/unsloth Finetune框架

强化学习训练框架#

• [54.7k] hiyouga/LLaMA-Factory
• [14.7k] huggingface/trl
• [11.3k] volcengine/verl 字节
• [8.8k] modelscope/ms-swift
• [7.4k] OpenRLHF/OpenRLHF
• [1.5k] alibaba/ROLL

分布式训练#

• [2024.10] Liger Kernel: Efficient Triton Kernels for LLM Training
  • linkedin/Liger-Kernel
• [2023.04] PyTorch FSDP: Experiences on Scaling Fully Sharded Data Parallel
• [2019.10] ZeRO: Memory Optimizations Toward Training Trillion Parameter Models
  • [LLM]大模型显存计算公式与优化