MegaTrain: Full Precision Training of 100B+ Parameter Large Language Models on a Single GPU
This enables efficient large-scale model training on limited hardware, reducing costs and accessibility barriers for researchers and practitioners.
The paper tackles the problem of training large language models with over 100 billion parameters at full precision on a single GPU by introducing MegaTrain, a memory-centric system that stores parameters in host memory and uses optimizations like pipelined double-buffering and stateless layer templates, achieving reliable training of up to 120B parameters and 1.84x higher throughput than DeepSpeed ZeRO-3 for 14B models.
We present MegaTrain, a memory-centric system that efficiently trains 100B+ parameter large language models at full precision on a single GPU. Unlike traditional GPU-centric systems, MegaTrain stores parameters and optimizer states in host memory (CPU memory) and treats GPUs as transient compute engines. For each layer, we stream parameters in and compute gradients out, minimizing persistent device state. To battle the CPU-GPU bandwidth bottleneck, we adopt two key optimizations. 1) We introduce a pipelined double-buffered execution engine that overlaps parameter prefetching, computation, and gradient offloading across multiple CUDA streams, enabling continuous GPU execution. 2) We replace persistent autograd graphs with stateless layer templates, binding weights dynamically as they stream in, eliminating persistent graph metadata while providing flexibility in scheduling. On a single H200 GPU with 1.5TB host memory, MegaTrain reliably trains models up to 120B parameters. It also achieves 1.84$\times$ the training throughput of DeepSpeed ZeRO-3 with CPU offloading when training 14B models. MegaTrain also enables 7B model training with 512k token context on a single GH200.