1-bit Adam: Communication Efficient Large-Scale Training with Adam's Convergence Speed

Abstract

Scalable training of large models (like BERT and GPT-3) requires careful optimization rooted in model design, architecture, and system capabilities. From a system standpoint, communication has become a major bottleneck, especially on commodity systems with standard TCP interconnects that offer limited network bandwidth. Communication compression is an important technique to reduce training time on such systems. One of the most effective methods is error-compensated compression, which offers robust convergence speed even under 1-bit compression. However, state-of-the-art error compensation techniques only work with basic optimizers like SGD and momentum SGD, which are linearly dependent on the gradients. They do not work with non-linear gradient-based optimizers like Adam, which offer state-of-the-art convergence efficiency and accuracy for models like BERT. In this paper, we propose 1-bit Adam that reduces the communication volume by up to 5×, offers much better scalability, and provides the same convergence speed as uncompressed Adam. Our key finding is that Adam's variance (non-linear term) becomes stable (after a warmup phase) and can be used as a fixed precondition for the rest of the training (compression phase). Experiments on up to 256 GPUs show that 1-bit Adam enables up to 3.3× higher throughput for BERT-Large pre-training and up to 2.9× higher throughput for SQuAD fine-tuning. In addition, we provide theoretical analysis for our proposed work.

Maximizing Communication Efficiency for Large-scale Training via 0/1 Adam

“…Bernstein et al (2018b); Sohn et al (2019); Le Phong & Phuong (2020); Lyu (2021) investigate the robustness of 1-bit SGD. Perhaps the closest works to this paper are (Tang et al, 2021;Li et al, 2021), which propose using two-stage training to enable 1-bit Adam and 1-bit Lamb, respectively. Among all the variants of 1-bit communication, the design with error feedback mechanism has shown to work best both empirically (Seide et al, 2014) and theoretically (Karimireddy et al, 2019).…”

“…Limitations of the state-of-the-art 1-bit Adam. Tang et al (2021) undertook the first investigation of this question and proposed 1-bit Adam. The algorithm follows a two-stage training paradigm: first run Adam with full-precision communication (full-precision stage 1 ); and then switch to 1 bit when the variance state, i.e.…”

“…the running average of second moment gradient in Adam, becomes stable (compression stage). While this paradigm allows drastic data volume reduction over original Adam, the experimental results from (Tang et al, 2021) indicate the full-precision stage still incurs non-trivial overhead. For instance, it is shown that when training the BERT-Large model, 1-bit Adam sends each parameter with 5.69 bits on average, which is not close to 1 bit as we would expect.…”

BAGUA: Scaling up Distributed Learning with System Relaxations

“…These two design decisions enable the flexibility and efficiency of BAGUA-to implement a new advanced algorithm with system relaxation (e.g., 1-big Adam [79] or Decentralized SGD [15]), in BAGUA, a developer does not need to worry about manually balancing communications with computations; instead, she can specify, at a high-level, the logical semantics and BAGUA will automatically optimize its execution. In this section, we first provide a high-level system overview, followed by a descriptions of these primitives and their implementations, and then the simple, but effective, optimization framework in BAGUA.…”

“…QSGD [4], a quantized (8-bit) DP-SG algorithm, implemented with C LP S primitive without error compensation. 1-bit Adam [79], a quantized (1-bit) distributed learning algorithm, implemented with by C LP S primitive with error compensation. Decen-32bits, a decentralized training algorithm with the random probing method to exchange the model parameters in each iteration, implemented with D FP S. Decen-8bits [17], a ring-based decentralized training algorithm with quantization, implemented with D LP S. Async, asynchronous centralized DP-SG.…”