Quantized distributed training of large models with convergence guarantees

conference paper Quantized distributed training of large models with convergence guarantees PMLR published yes Ilia Markov author D0CF4148-C985-11E9-8066-0BDEE5697425 Adrian Vladu author Qi Guo author Dan-Adrian Alistarh author 4A899BFC-F248-11E8-B48F-1D18A9856A870000-0003-3650-940X DaAl department ICML: International Conference on Machine Learning Elastic Coordination for Scalable Machine Learning project Communication-reduction techniques are a popular way to improve scalability in data-parallel training of deep neural networks (DNNs). The recent emergence of large language models such as GPT has created the need for new approaches to exploit data-parallelism. Among these, fully-sharded data parallel (FSDP) training is highly popular, yet it still encounters scalability bottlenecks. One reason is that applying compression techniques to FSDP is challenging: as the vast majority of the communication involves the model’s weights, direct compression alters convergence and leads to accuracy loss. We present QSDP, a variant of FSDP which supports both gradient and weight quantization with theoretical guarantees, is simple to implement and has essentially no overheads. To derive QSDP we prove that a natural modification of SGD achieves convergence even when we only maintain quantized weights, and thus the domain over which we train consists of quantized points and is, therefore, highly non-convex. We validate this approach by training GPT-family models with up to 1.3 billion parameters on a multi-node cluster. Experiments show that QSDP preserves model accuracy, while completely removing the communication bottlenecks of FSDP, providing end-to-end speedups of up to 2.2x. ML Research Press2023Honolulu, Hawaii, HI, United States eng Proceedings of the 40th International Conference on Machine Learning 2640-3498 2302.02390 20224020-24044 https://research-explorer.ista.ac.at/record/17490 Markov, I., Vladu, A., Guo, Q., & Alistarh, D.-A. (2023). Quantized distributed training of large models with convergence guarantees. In Proceedings of the 40th International Conference on Machine Learning (Vol. 202, pp. 24020–24044). Honolulu, Hawaii, HI, United States: ML Research Press. Markov, Ilia, et al. “Quantized Distributed Training of Large Models with Convergence Guarantees.” Proceedings of the 40th International Conference on Machine Learning, vol. 202, ML Research Press, 2023, pp. 24020–44. Markov, Ilia, Adrian Vladu, Qi Guo, and Dan-Adrian Alistarh. “Quantized Distributed Training of Large Models with Convergence Guarantees.” In Proceedings of the 40th International Conference on Machine Learning, 202:24020–44. ML Research Press, 2023. I. Markov, A. Vladu, Q. Guo, and D.-A. Alistarh, “Quantized distributed training of large models with convergence guarantees,” in Proceedings of the 40th International Conference on Machine Learning, Honolulu, Hawaii, HI, United States, 2023, vol. 202, pp. 24020–24044. Markov I, Vladu A, Guo Q, Alistarh D-A. 2023. Quantized distributed training of large models with convergence guarantees. Proceedings of the 40th International Conference on Machine Learning. ICML: International Conference on Machine Learning, PMLR, vol. 202, 24020–24044. Markov I, Vladu A, Guo Q, Alistarh D-A. Quantized distributed training of large models with convergence guarantees. In: Proceedings of the 40th International Conference on Machine Learning. Vol 202. ML Research Press; 2023:24020-24044. I. Markov, A. Vladu, Q. Guo, D.-A. Alistarh, in:, Proceedings of the 40th International Conference on Machine Learning, ML Research Press, 2023, pp. 24020–24044. 144612023-10-29T23:01:17Z2026-10-02T11:20:05Z