--- title: "Megatron-LM" created: 2026-06-24 updated: 2026-06-24 type: concept tags: ["training-framework", "distributed-training", "model-parallelism"] sources: - "[[unlimited-ocr-works-2026]]" --- # Megatron-LM Megatron-LM(NVIDIA, arXiv:1909.08053)是训练数十亿参数大语言模型的分布式训练框架,提供张量并行、流水线并行等模型并行策略。Unlimited OCR 的训练 pipeline 基于 Megatron-LM 构建。 ## 参考 - [[unlimited-ocr-works-2026]]