MEMO: Fine-grained Tensor Management For Ultra-long Context LLM Training

TL;DR

MEMO introduces a fine-grained activation memory management framework that enables efficient training of ultra-long context LLMs by offloading activations to CPU memory and optimizing memory reuse, significantly improving GPU utilization.

Contribution

The paper presents MEMO, a novel framework that manages activation memory at a fine granularity, reducing fragmentation and communication overhead for long context LLM training.

Findings

Achieves 1.97x and 1.80x MFU improvements over Megatron-LM and DeepSpeed.

Enables training of 7B LLM with 1 million sequence length on 8 GPUs.

Reduces memory fragmentation and recomputation overhead.

Abstract

Nowadays, Large Language Models (LLMs) have been trained using extended context lengths to foster more creative applications. However, long context training poses great challenges considering the constraint of GPU memory. It not only leads to substantial activation memory consumption during training, but also incurs considerable memory fragmentation. To facilitate long context training, existing frameworks have adopted strategies such as recomputation and various forms of parallelisms. Nevertheless, these techniques rely on redundant computation or extensive communication, resulting in low Model FLOPS Utilization (MFU). In this paper, we propose MEMO, a novel LLM training framework designed for fine-grained activation memory management. Given the quadratic scaling of computation and linear scaling of memory with sequence lengths when using FlashAttention, we offload memory-consuming activations to CPU memory after each layer's forward pass and fetch them during the backward pass. To maximize the swapping of activations without hindering computation, and to avoid exhausting limited CPU memory, we implement a token-wise activation recomputation and swapping mechanism. Furthermore, we tackle the memory fragmentation issue by employing a bi-level Mixed Integer Programming (MIP) approach, optimizing memory reuse across transformer layers. Empirical results demonstrate that MEMO achieves an average of 1.97x and 1.80x MFU compared to Megatron-LM and DeepSpeed, respectively. This improvement is attributed to MEMO's ability to minimize memory fragmentation, reduce recomputation and intensive communication, and circumvent the delays associated with the memory reorganization process due to fragmentation. By leveraging fine-grained activation memory management, MEMO facilitates efficient training of 7B LLM with 1 million sequence length on just 8 A800 GPUs, achieving an MFU of 52.30%.