PapersMulti-sourceSelected

EAServe Boosts Multimodal LLM Serving Goodput by 4.3x

EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models

Why we picked it

Disaggregated serving mostly assumes a two-stage text pipeline, so adding Encode as a regulated control point changes how multimodal workloads should be scheduled and provisioned. Engineers running MLLM inference at scale — whether with vLLM or NVIDIA Dynamo — can compare EAServe's EPD design and its automated configuration search against their current setups when encode GPUs sit idle while prefill and decode starve. The 4.3x and 1.7x goodput figures give a concrete yardstick under identical SLOs.

EAServe is a serving system for multimodal LLMs that treats the Encode stage as the control point of the Encode-Prefill-Decode (EPD) pipeline.

Body

Its runtime applies load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning, while its Hybrid Auto Selection layer searches GPU allocation, encode batch size, and offload ratio via TPE-based Bayesian optimization.

On three MLLM architectures covering image, video, and audio, EAServe achieves up to 4.3x higher goodput than NVIDIA Dynamo and 1.

7x higher than vLLM under identical SLO constraints.It also sustains more balanced GPU utilization across the pipeline and finds near-optimal configurations faster than baseline search methods.

Original · arXiv cs.LG

EAServe: Encode-Aware Disaggregated Serving for Multimodal Large Language Models

Disaggregating the two stages, Prefill and Decode, onto separate GPU pools is now a standard optimization for (text-only) LLM serving. However, multimodal LLMs (MLLMs), which add a third phase, Encode, pose new challenges for resource allocation. Encode turns images, video, or audio into embeddings that the language model can consume, yielding a three-stage Encode-Prefill-Decode (EPD) pipeline. Existing frameworks offer only partial answers: text-only PD systems lack Encode, while EPD frameworks expose it as a separate service without regulating downstream request flow. The pipeline also carries a structural resource imbalance: every request enters through Encode before downstream work can begin, yet per-request execution leaves the encode GPU severely underutilized even at high loads, starving the downstream Prefill and Decode workers. Addressing this, we reposition Encode as the control point of the EPD pipeline, exposing three tightly coupled dimensions: when work enters downstream, where prefill executes, and how the GPU is shared. We instantiate this in EAServe across two co-designed layers. Its runtime manages load-adaptive micro-batching, rate-controlled partial offload to a co-resident prefill worker, and dynamic SM partitioning for predictable co-location. The configuration layer, Hybrid Auto Selection (HAS), navigates the joint space of GPU allocation, encode batch size, and offload ratio by pruning unbalanced allocations with per-stage capacity profiling and refining the remainder through TPE-based Bayesian optimization. Evaluated on three MLLM architectures spanning image, video, and audio, EAServe delivers up to 4.3x and 1.7x higher goodput than NVIDIA Dynamo and vLLM, respectively, under identical SLO constraints, sustains more balanced and higher GPU utilization across the EPD pipeline, and reaches near-optimal configurations faster than baseline search methods.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include EAServe, vLLM, NVIDIA Dynamo, multimodal.

Source attribution: curated from arXiv cs.LG (@Kunxiong Zhu, Zhihao Shu, Hangyu Zheng, Minghai Qin, Miao Yin, Gagan Agrawal, Wei Niu). Cross-source corroboration: 6 independent sources have covered this development (including DAIR.AI, Genspark, rohanpaul_ai), confirming industry-wide signal strength.