Scaling State-Space Models on Multiple GPUs with Tensor Parallelism

Selective state space models (SSMs) have rapidly become a compelling backbone for large language models, especially for long-context workloads. Yet in deployment, their inference performance is often bounded by the memory capacity, bandwidth, and latency limits of a single GPU, making multi-GPU execution increasingly necessary. Although tensor parallelism (TP) is widely used to scale Transformer inference, applying it to selective SSM blocks is non-trivial because the SSM mixer couples large projections with a sequence-wise recurrent state update and local mixing whose efficiency depends on preserving locality and avoiding synchronization in the critical path. This paper presents a communication-efficient TP design for selective SSM inference that addresses three practical engineering challenges: enabling TTFT improvements via an SSM state cache across prefill and decode, partitioning the mixer's packed parameter tensor so that recurrent updates remain local while minimizing communication, and reducing TP aggregation overhead with quantized AllReduce. We evaluate on three representative SSM-based LLMs spanning pure-SSM and hybrid architectures - Mamba, Falcon-Mamba, and Zamba - on NVIDIA A6000 and A100 clusters. Our experiments show substantial throughput gains from tensor-parallel SSM inference, improving batch-request throughput by ~1.6-2.1x on 2 GPUs and ~2.6-4.0x on 4 GPUs for Mamba, with the largest benefits at long context lengths, and achieving a further ~10-18% throughput improvement from quantized all-reduce by lowering synchronization bandwidth overhead.

Representation Similarity: A Better Guidance of DNN Layer Sharing for Edge Computing without Training

Edge computing has emerged as an alternative to reduce transmission and processing delay and preserve privacy of the video streams. However, the ever-increasing complexity of Deep Neural Networks (DNNs) used in video-based applications (e.g. object detection) exerts pressure on memory-constrained edge devices. Model merging is proposed to reduce the DNNs' memory footprint by keeping only one copy of merged layers' weights in memory. In existing model merging techniques, (i) only architecturally identical layers can be shared; (ii) requires computationally expensive retraining in the cloud; (iii) assumes the availability of ground truth for retraining. The re-evaluation of a merged model's performance, however, requires a validation dataset with ground truth, typically runs at the cloud. Common metrics to guide the selection of shared layers include the size or computational cost of shared layers or representation size. We propose a new model merging scheme by sharing representations (i.e., outputs of layers) at the edge, guided by representation similarity S. We show that S is extremely highly correlated with merged model's accuracy with Pearson Correlation Coefficient |r| > 0.94 than other metrics, demonstrating that representation similarity can serve as a strong validation accuracy indicator without ground truth. We present our preliminary results of the newly proposed model merging scheme with identified challenges, demonstrating a promising research future direction.