Large language models in IoT–edge–cloud settings face bursty, heterogeneous requests that make pipeline-parallel inference prone to micro-batch imbalance and communication stalls, causing GPU idle time and SLO violations. We propose a runtime-adaptive scheduler that jointly tunes token budgets and micro-batch counts to balance prefill/decode workloads and minimize pipeline bubbles under changing compute and network conditions. On a four-node pipeline-parallel cluster across Llama-2-13b and Qwen2.5-14b at 100/1000 Mbps, our method outperforms vLLM and SGLang, reducing GPU idle time by up to 55% and improving throughput by up to 1.61 × while improving TTFT/ITL SLO satisfaction. These results show that dynamic scheduling is essential for scalable, latency-stable LLM inference in IoT–edge–cloud environments.

As large language models (LLMs) pursue higher accuracy, their model sizes have surged, substantially increasing GPU memory consumption. Prior work mitigates this issue by distributing the memory burden across multiple GPUs. However, on clusters interconnected via Ethernet, the resulting computational intensity is insufficient to hide the significant network latency. Achieving a favorable compute-to-communication ratio is further constrained by the memory required to cache the massive activations generated during the forward pass. PyAO, proposed in this paper, effectively offloads activations, selects offloading strategies based on their offloading efficiency, and minimizes data-movement bottlenecks, thereby enabling larger micro-batch sizes. In Ethernet-interconnected cluster environments, experiments on popular models—including OPT-1.3B, GPT-0.8B, and Llama-1.2B—demonstrate that PyAO reduces peak GPU memory by up to 1.94× at the same micro-batch size, enables up to 2.5× larger batch sizes, and accelerates training by up to 3.63× relative to the baseline.