Distributed Lion: Scalable Sign-Based Optimization

• Distributed Lion is a distributed optimization algorithm that extends the sign-based Lion optimizer to reduce communication overhead in large-scale deep learning.
• It aggregates binary updates via Majority Vote or Averaging, achieving provable convergence and matching statistical efficiency with full-precision methods.
• Empirical evaluations demonstrate that Distributed Lion maintains high accuracy across vision and language tasks even in bandwidth-constrained settings.

Distributed Lion refers to a class of distributed optimization algorithms that generalize the single-worker Lion optimizer—a first-order, sign-based momentum method—into scalable, communication-efficient distributed settings. These algorithms leverage the sign operation central to Lion to minimize per-iteration communication, enabling distributed training of large-scale deep models with significant bandwidth reduction and provable convergence properties, while matching or exceeding the statistical efficiency of standard aggregation-based optimizers.

1. Formal Algorithmic Structure

Distributed Lion operates within a parameter-server or similar synchronous distributed framework. The canonical formulation comprises $N$ workers and a central server. Each worker $i$ maintains its own set of local parameters $x_{i,t}$ and momentum buffer $m_{i,t}$, updating these at each round as follows:

• Worker step:

1. Sample data $\xi_{i,t}$, compute gradient $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$.
2. Update momentum: $m_{i,t+1} = \beta_2 m_{i,t} + (1-\beta_2) g_{i,t}$.
3. Compute update: $$d_{i,t} = \operatorname{sign}\left(\beta_1 m_{i,t} + (1-\beta_1) g_{i,t}\right)\in\{-1,+1\}^d$$.
4. Transmit $d_{i,t}$ (a binary vector) to the server.

• Server aggregation:
  • Majority Vote (MaVo): $i$0.
  • Averaging (Avg): $i$1.
  • 3. Broadcast $i$2 to all workers.
• Parameter update:

$i$3.

This scheme preserves the sign-based nature of Lion updates while requiring only binary or low-bitwidth communications per iteration (Liu et al., 2024). Key variants include Majority Vote (strictly $i$4 bit per parameter per direction) and Averaging (using $i$5 bits per parameter per direction).

2. Theoretical Guarantees and Constrained Optimization Perspective

Distributed Lion inherits and extends the constrained optimization interpretation of the original Lion. The weight-decay operation implicitly enforces an $i$6-box constraint, so

$i$7

The dynamics exhibit two phases:

• Phase I: Rapid contraction of parameters into the constraint set, with the box distance decaying exponentially [(Liu et al., 2024), Prop. A.5].
• Phase II: Optimization within the $i$8 box, converging to KKT stationarity as measured by

$i$9

Convergence rates depend on the aggregation method:

• Majority Vote (MaVo): Per-iteration expected KKT residual decays as $x_{i,t}$0.
• Averaging (Avg): Similar, but the variance term persists with $x_{i,t}$1 [(Liu et al., 2024), Theorem 4.6/4.8].
• In centralized or full-precision distributed settings, the rate matches $x_{i,t}$2 for sign-based methods (Jiang et al., 17 Aug 2025).

Bandlimited variants, using unbiased sign compression for both upward and downward communication, achieve provably controlled increases in asymptotic rates, e.g., $x_{i,t}$3 for the most communication-efficient version (Jiang et al., 17 Aug 2025).

3. Communication Complexity and Compression

Distributed Lion methods achieve marked reductions in per-iteration communication, tabulated as follows (Liu et al., 2024):

Method	Worker→Server	Server→Worker
Global Lion/AdamW	$x_{i,t}$4 bits	$x_{i,t}$5 bits
TernGrad	$x_{i,t}$6 bits	$x_{i,t}$7 bits
Deep Grad. Compress (DGC)	$x_{i,t}$8 bits	$x_{i,t}$9 bits
Distributed Lion-Avg	$m_{i,t}$0 bit	$m_{i,t}$1 bits
Distributed Lion-MaVo	$m_{i,t}$2 bit	$m_{i,t}$3 bit

This results in $m_{i,t}$4 reduction (MaVo) or $m_{i,t}$5–$m_{i,t}$6 reduction (Avg, $m_{i,t}$7) per iteration compared to full-precision approaches. The method is orthogonal to existing sparsification and quantization strategies and can be hybridized for further savings.

4. Empirical Performance and Applicability

Distributed Lion demonstrates strong empirical results across diverse deep learning settings:

• Vision (CIFAR-10, ImageNet-1K): On ViT-Small and ViT-B/16 models, test accuracy with Distributed Lion MaVo is within $m_{i,t}$8 of full-precision Lion or AdamW, even with up to $m_{i,t}$9 workers (Liu et al., 2024).
• Language (GPT2++, LLaMA-7B): Perplexity and few-shot tuning performance match or slightly exceed full-precision baselines. Communication-efficient Lion variants consistently outperform TernGrad, DGC, and other sign-based compression approaches for both accuracy and bandwidth trade-off.
• Batch size and scalability: As the number of workers increases, all methods experience minor accuracy decrements (due to batch noise reduction), but Distributed Lion variants maintain competitive statistical efficiency.

Applicability is particularly strong in scenarios where:

• Network bandwidth is limited (e.g., multi-site, wireless).
• Models are large enough that communication is the bottleneck.
• High-frequency, low-precision updates are acceptable or desirable.

5. Advanced Extensions: Momentum Synchronization and Quantization

Extended variants (e.g., Lion Cub (Ishikawa et al., 2024)) further address the communication bottleneck by combining:

• Custom collectives: Efficient 1-bit or $\xi_{i,t}$0-bit allreduce strategies, including fused bit-packing and direct CUDA/NCCL implementations, tuned for high-latency and bandwidth-constrained networks.
• Quantization: Standard sign-quantization ($\xi_{i,t}$1) and novel $\xi_{i,t}$2-scale quantizers for few-bit encoding, empirically matching full-precision Lion's updates over $\xi_{i,t}$3 of the time.
• Momentum synchronization: Selective or periodic momentum buffer averaging (e.g., every $\xi_{i,t}$4 steps for selected layers) is required for certain hyperparameter regimes, particularly when momentum decay rates are lower.

Empirically, these techniques enable up to $\xi_{i,t}$5 reduction in end-to-end training time on Ethernet-based clusters, without sacrificing final model quality.

6. Distributed Lion in Federated and Heterogeneous Settings

The canonical Lion update naturally extends to federated optimization (FedLion (Tang et al., 2024)). In FedLion, clients perform local sign-based Lion steps with momentum, uploading quantized integer vectors and optionally momentum buffers. Compared to FedAvg, FedLion:

• Achieves a per-round uplink close to $\xi_{i,t}$6 bits (plus $\xi_{i,t}$7 bits per parameter for the quantized update, $\xi_{i,t}$8 = local epochs/steps).
• Requires $\xi_{i,t}$9 the rounds to reach the same accuracy, compared to state-of-the-art adaptive federated algorithms.

Convergence is established under standard bounded-variance, smoothness, and system heterogeneity assumptions, with an $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$0-rate in squared $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$1-norm, outperforming the $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$2-rate of FedAvg in dense gradient regimes.

7. Practical Trade-offs and Limitations

The choice of Majority Vote vs Averaging impacts the communication/accuracy trade-off:

• Majority Vote: Strict $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$3-bit exchange, robust to high noise, preferable for small $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$4 and noisy updates.
• Averaging: Slightly higher communication cost, potentially better accuracy, particularly at large $g_{i,t} = \nabla f(x_{i,t}; \xi_{i,t})$5 where batch-level gradient noise is reduced.

Distributed Lion assumes that local momentum/parameter drift due to sign-only communication can be effectively controlled by periodic synchronization or rich quantization when necessary, but for some tasks and optimizer configurations (e.g., low momentum rates), additional synchronization may be required (Ishikawa et al., 2024).

A plausible implication is that Distributed Lion sets a practical lower bound on communication in modern distributed deep learning and forms a basis for hybrid methods combining sign compression, gradient sparsification, or error compensation. Convergence, scalability, and statistical efficiency have been established rigorously and validated empirically across vision and language benchmarks (Liu et al., 2024, Jiang et al., 17 Aug 2025).

---

References

• "Communication Efficient Distributed Training with Distributed Lion" (Liu et al., 2024)
• "Lion Cub: Minimizing Communication Overhead in Distributed Lion" (Ishikawa et al., 2024)
• "FedLion: Faster Adaptive Federated Optimization with Fewer Communication" (Tang et al., 2024)
• "Convergence Analysis of the Lion Optimizer in Centralized and Distributed Settings" (Jiang et al., 17 Aug 2025)