Cost-Efficient Multimodal LLM Inference via Cross-Tier GPU Heterogeneity

TL;DR

This paper introduces a cost-effective approach for multimodal large language model inference by leveraging cross-tier GPU heterogeneity, optimizing partitioning to reduce data transfer and improve efficiency on commodity hardware.

Contribution

It identifies the optimal modality boundary for partitioning, develops a phase-aware runtime, and demonstrates significant cost and performance improvements with heterogeneous GPU deployment.

Findings

Partitioning at the modality boundary minimizes cross-device transfer.

Heterogeneous deployment achieves up to 40.6% cost savings.

HeteroServe improves throughput and token efficiency on real models.

Abstract

Multimodal large language model (MLLM) inference splits into two phases with opposing hardware demands: vision encoding is compute-bound, while language generation is memory-bandwidth-bound. We show that under standard transformer KV caching, the modality boundary (between vision encoder and language model) minimizes cross-device transfer among all partition points that preserve standard stage-based execution. Partitioning here reduces transfer complexity from $O(L * s_ctx)$ bytes (GB-scale KV caches under stage-level disaggregation) to $O(N_v * d)$ bytes (MB-scale embeddings), an O(L) reduction where L is the transformer depth. The result holds across attention mechanisms (MHA/GQA), dynamic vision resolutions, and model scales, and the advantage grows as models deepen. A direct implication is that existing stage-level disaggregation systems are constrained to high-bandwidth interconnects (e.g., NVLink), whereas modality-level disaggregation enables cross-tier heterogeneous serving over commodity PCIe. A closed-form cost model shows that heterogeneous deployment is cost-optimal under phase-separable workloads (predicts 31.4% savings; observed 40.6%). We build HeteroServe, a phase-aware runtime with modality-level partitioning and cross-tier scheduling, and evaluate it on LLaVA-1.5-7B and Qwen2.5-VL against vLLM v0.3.0. On identical 4xA100 hardware, engine optimizations raise throughput by up to 54%. Under a fixed budget, a heterogeneous cluster (\$38k) improves Tokens/\$ by 37% over a homogeneous baseline (\$64k) without degrading latency.